Chapter 3

Stochastic Time Series Methods

Abstract Hydrology was mainly dominated by deterministic approaches until the

mid-twentieth century. However, the deterministic approaches suffered from our
lack of knowledge on the exact nature of hydrologic system dynamics and, hence,
the exact governing equations required for models. This led to the development and
application of stochastic methods in hydrology, which are based on the concepts of
probability and statistics. Since the 1950s–1960s, hydrology has witnessed the
development of a large number of stochastic time series methods and their appli-
cations. The existing stochastic methods can be broadly grouped into two categories:
parametric and nonparametric. In the parametric methods, the structure of the models
is defined a priori and the number and nature of the parameters are generally fixed in
advance. On the other hand, the nonparametric methods make no prior assumptions
on the model structure, and it is essentially determined from the data themselves.
This chapter presents an overview of stochastic time series methods in hydrology.
First, a brief account of the history of development of stochastic methods is pre-
sented. Next, the concept of time series and relevant statistical characteristics and
estimators are described. Finally, several popular parametric and nonparametric
methods and their hydrologic applications are discussed.

3.1 Introduction

The complex, irregular, and random-looking nature of hydrologic time series,

combined with our lack of knowledge of the exact governing equations required in
deterministic models, motivated the development and applications of methods
based on probability and statistics for modeling and prediction of such time series.
In this regard, stochastic methods are an appropriate means, since they aim at
prediction (or estimation) of data in a probabilistic manner, with particular emphasis
on the statistical characteristics of the data (e.g. mean, standard deviation, variance)
and with proper consideration to uncertainty in such predictions.
Stochastic time series methods in hydrology has a long and rich history. The
emergence of stochastic methods during the first half of the 20th century (e.g. Lévy

© Springer Science+Business Media Dordrecht 2017 63

B. Sivakumar, Chaos in Hydrology, DOI 10.1007/978-90-481-2552-4_3
64 3 Stochastic Time Series Methods

1925, 1948; Doob 1934, 1938, 1940, 1945; Berndstein 1938; Itô 1944, 1946; Bochner
1949) led to their introduction in hydrology in the 1950s (Hurst 1951, 1956; Hannan
1955; Le Cam 1961), especially for studying storage in reservoirs and rainfall mod-
eling. Further advances in the development of many stochastic time series methods
around the middle of the 20th century (e.g. Wiener 1949; Feller 1950; Bartlett 1955;
Box and Jenkins 1970; Brillinger 1975) led to a real impetus in applying stochastic
methods in hydrology during 1960s–1970s (e.g. Thomas and Fiering 1962; Matalas
1963a, b, 1967; Yevjevich 1963, 1972; Roesner and Yevjevich 1966; Fiering 1967;
Harms and Campbell 1967; Quimpo 1967; Mandelbrot and Wallis 1968, 1969;
Carlson et al. 1970; Valencia and Schaake 1973; McKerchar and Delleur 1974; Haan
et al. 1976; Hipel et al. 1977; Hipel and McLeod 1978a, b; Lawrence and Kottegoda
1977; Lettenmaier and Burges 1977; Delleur and Kavvas 1978; Klemeš 1978; Hirsch
1979); see also Salas et al. (1995) for a comprehensive account.
Since then, stochastic methods in hydrology have witnessed tremendous
advances, both in terms of theoretical development and in terms of applications in
many different areas and problems in hydrology. Studies during the 1980s and early
1990s mainly focused on the development and applications of parametric methods
in hydrology (e.g. Gupta and Waymire 1981; Kavvas and Delleur 1981; Salas and
Smith 1981; Salas and Obeysekera 1982; Srikanthan and McMahon 1983; Rao and
Rao 1984; Rodriguez-Iturbe et al. 1987; Koutsoyiannis and Xanthopoulos 1990);
see also Bras and Rodriguez-Iturbe (1985), Gelhar (1993), and Salas (1993) for
some comprehensive accounts. However, the development of nonparametric
approaches (e.g. Efron 1979; Silverman 1986; Eubank 1988; Härdle and Bowman
1988; Efron and Tibishirani 1993) to overcome the difficulties associated with the
estimation of parameters in parametric methods led to their applications in
hydrology. Such applications have been gaining significant momentum during the
past two decades or so (e.g. Kendall and Dracup 1991; Lall 1995; Lall and Sharma
1996; Vogel and Shallcross 1996; Tarboton et al. 1998; Buishand and Brandsma
2001; Sharma and O’Neill 2002; Prairie et al. 2006; Mehrotra and Sharma 2010;
Salas and Lee 2010; Wilks 2010; Li and Singh 2014).
In more recent years, stochastic time series methods have been finding
increasing applications in the study of hydrologic extremes, parameter estimation in
hydrologic models, and downscaling of global climate model outputs, among others
(e.g. Fowler et al. 2007; Vrugt et al. 2008; Maraun et al. 2010; Mehrotra and
Sharma 2012; Pui et al. 2012; Grillakis et al. 2013; Bordoy and Burlando 2014;
D’Onofrio et al. 2014; Sikorska et al. 2015; Wasko et al. 2015; Langousis et al.
2016). Although many modern nonlinear approaches have found important places
in hydrology in recent times (see Chap. 4 for details), stochastic time series
methods continue to dominate, for various reasons, including for their great flexi-
bility. Extensive details of the applications of stochastic methods in hydrology are
available in Kottegoda (1980), Bras and Rodriguez-Iturbe (1985), Gelhar (1993),
Clarke (1994), Haan (1994), Hipel and McLeod (1994), Salas et al. (1995),
Govindaraju (2002), and McCuen (2003), among others.
This chapter presents an overview of stochastic time series methods in hydrol-
ogy. First, a brief account of the history of development and fundamentals of
3.1 Introduction 65

stochastic methods is presented. Next, the definition and classification of hydrologic

time series are provided, followed by a description of their relevant statistical
characteristics and estimators. Then, several parametric methods and nonparametric
methods that have found widespread applications in hydrology are described. The
chapter ends with a brief summary of stochastic methods in hydrology.
Due to the domination of stochastic methods for over half a century, the liter-
ature on stochastic methods in hydrology is enormous. Therefore, it is impossible to
cover all the available literature here. For additional details, the reader is directed to
the many studies cited in this chapter and the references therein. Furthermore, the
stochastic methods described in this chapter are those that either assume linearity in
the process or make no prior assumptions regarding linearity/nonlinearity. There are
indeed many stochastic methods that explicitly assume nonlinearity of the process.
A brief account of such nonlinear stochastic methods is presented in Chap. 4, along
with many other modern nonlinear time series methods.

3.2 Brief History of Development of Stochastic Methods

The term ‘stochastic’ was derived from the Greek word ‘Στόχος’ (stochos), meaning
‘target.’ However, in the context of modern science, ‘stochastic’ generally means
‘random’ or refers to the presence of randomness. Stochastic methods, in essence,
aim at predicting (or estimating) the value of some variable at non-observed times or
at non-observed locations in a probabilistic manner, while also stating how uncertain
the predictions are. The methods place emphasis on the statistical characteristics
(e.g. mean, standard deviation, variance) of relevant processes.
Probabilistic approaches with emphasis on statistical characteristics of data
observed in nature have a very long history. However, stochastic models, as they are
seen in their current form, originated around the mid-twentieth century; see Wiener
(1949) and Feller (1950) for some early accounts. Such developments were based on
many earlier developments, including stochastic differential equations, stochastic
integrals, Markov chains, Brownian motion, and diffusion equations (e.g. Berndstein
1938; Doob 1938, 1940, 1945; Lévy 1937, 1948; Itô 1944, 1946; Mann 1945;
Bochner 1949). The 1950s–1960s witnessed some important developments, includ-
ing advances in diffusion processes, harmonic analysis, Markov processes, branching
processes, and random walk theory; see Doob (1953), Bartlett (1955), Bochner
(1955), Loève (1955), Dynkin (1959), Spitzer (1964), and McKean (1969) for details.
A major shift in stochastic methods and their applications occurred around the
1970s. This could mainly be attributed to the work of Box and Jenkins (1970), who
presented an exhaustive account of stochastic time series methods until then, and
several others (e.g. Jazwinski 1970; Brillinger 1975; Karlin and Taylor 1975). The
methods presented in Box and Jenkins (1970) have and continue to be widely
followed in various scientific fields, including hydrology. Some of the models that
have found widespread applications are the autoregressive (AR) models, moving
average (MA) models, autoregressive moving average (ARMA) models, Markov
66 3 Stochastic Time Series Methods

chain models, point process models, and their variants. Parallel theoretical devel-
opments in the concepts of fractional Brownian motion (fBm), fractional Gaussian
noise (fGn), Lévy processes, and broken line processes (e.g. Mandelbrot and Van
Ness 1968; Mandlbrot and Wallis 1969; Mandelbrot 1972) also resulted in many
different types of models and subsequently found numerous applications in different
fields. Indeed, this period also witnessed a significant shift in the study of
scale-invariance or self-similarity of natural processes, especially with the works of
Mandelbrot and with the coining of the term ‘fractal’ (e.g. Mandelbrot 1967, 1977);
see also Mandelbrot (1983) for an exhaustive coverage.
These advances further propelled the development of stochastic methods and
their applications in different fields; see Chatfield (1989), Brockwell and Davis
(1991), Ross (1996), and Gardiner (2009) for details. However, due to the mainly
parametric nature of most of these methods, the models generally suffered from
certain important limitations, such as: (1) since the structure of the models is
defined a priori and the number and nature of the parameters are generally fixed in
advance, there is little flexibility; (2) the parameter estimation procedure is often
complicated, especially when the number of parameters is large; and (3) the models
are often not able to capture several important properties of the time series. This led
to the proposal of many nonparametric approaches since the 1980s (Efron 1979;
Silverman 1986; Eubank 1988; Härdle and Bowman 1988) and resulted in another
major shift in the development and applications of stochastic time series methods.
The nonparametric approaches make no prior assumptions on the model structure.
Instead, the model structure is determined from the data themselves. Although
nonparametric methods may also involve parameters, the number and nature of the
parameters are not fixed in advance and, thus, are flexible. Since their emergence,
many nonparametric methods have been developed, including those based on
kernel density estimate, block bootstrap, k-nearest neighbor bootstrap, k-nearest
neighbor with local polynomial, hybrid models, and others. Comprehensive
accounts of the concepts and applications of nonparametric methods can be found
in Efron and Tibishirani (1993), Higgins (2003), Sprent and Smeeton (2007), and
Hollander et al. (2013) among others.

3.3 Hydrologic Time Series and Classification

A hydrologic time series is a set of observations of a hydrologic variable (e.g.

rainfall, streamflow) made at a particular location over time (i). A hydrologic time
series can be either continuous or discrete. For instance, a streamflow monitoring
device at a location in a river provides a continuous record of river stage and
discharge X(i) through time. A plot of the flow hydrograph X(i) versus time i is a
continuous time series. Sampling this continuous time series at discrete points in
time or integrating it over successive time intervals results in a discrete time series,
denoted by Xi. Hydrologic time series are generally measured and studied at discrete
timescales (e.g. hourly, daily, monthly, annual). Such a discrete time series may be
3.3 Hydrologic Time Series and Classification 67

denoted as Xi, i = 1, 2, …, N, where N is the total length of the observed time and,
thus, often represents the total number of data (or points) in the time series. As the
study of discrete time series is the widely used practice in hydrology, the presen-
tation in this chapter (and in the rest of this book) focuses on such a time series only.
In addition to continuous and discrete time series, hydrologic time series may be
further classified into different categories depending upon the nature of the obser-
vations and the general/specific properties. A single or scalar or univariate time
series is a time series of a hydrologic variable at a given location. A multiple or
multivariate time series may represent either time series measured at more than one
location or time series of multiple variables at one location (or more). In a single time
series, if the value of a variable at time i, i.e. Xi, depends on the value of the variable
at other (earlier) times i–1, i–2, …, then the time series is called autocorrelated,
serially correlated or correlated in time; otherwise, it is uncorrelated. A similar
explanation also goes in a spatial sense. For instance, considering a multiple time
series (e.g. rainfall, streamflow), if the value of a variable (e.g. streamflow) at time
i at one location, Xi, depends on the value of another variable (e.g. rainfall) at the
same location (Yi) (or on the value of the same variable at another location) at times i,
i–1, i–2,…, then the two time series are cross-correlated. Autocorrelation and
cross-correlation in hydrologic time series can be simple or complex.
A seasonal time series is a series that corresponds to a time interval that is a fraction
of a year, such as muliples of a month. Since seasonality is an intrinsic property of
hydrologic systems, it is often useful, and even necessary, to study the seasonal
aspects of hydrologic systems. An intermittent time series is one when the variable of
interest takes on non-zero and zero values throughout the length of the time series.
Rainfall measured at finer timescales (e.g. hourly or daily) is often an excellent
example of an intermittent time series. A stationary time series is a time series that is
free of trends, shifts, or periodicity, implying that the statistical parameters of the
series (e.g. mean, variance) remain constant through time. If such parameters do not
remain constant through time, then the time series is called a nonstationary time series.
If the measurements are available at a regularly spaced interval, then the time series is
called a regularly spaced or regular time series. If the data are at irregular intervals,
then the time series is called an irregularly spaced or irregular time series.
Depending upon the properties or components, a time series can be partitioned
or decomposed into its component series. Some of the generic components of
hydrologic time series are trend, shift, and seasonality (in mean and variance as well
as in correlation). These components are often removed before employing
stochastic time series methods.

3.4 Relevant Statistical Characteristics and Estimators

Hydrologic systems and, hence, the observed hydrologic data exhibit a wide range
of characteristics. Chapter 2 presented an overview of many of the salient char-
acteristics, including complexity, correlation, trend, periodicity, cyclicity,
68 3 Stochastic Time Series Methods

seasonality, intermittency, stationarity, nonstationarity, linearity, nonlinearity,

determinism, randomness, scale and scale-invariance, self-organization and
self-organized criticality, threshold, emergence, feedback, and sensitivity to initial
conditions. Some of these characteristics have found particular significance in the
development and application of stochastic time series methods in hydrology. This
section briefly describes some of these characteristics and estimators, especially
those that have generally formed the basis for model development and for model
evaluation.
The most commonly used statistical properties for analyzing stationary or
nonstationary hydrologic time series are the sample mean, variance, coefficient of
variation, skewness coefficient, lag-τ autocorrelation function, and the power
spectrum. These are briefly described here.

3.4.1 Mean

The mean is the fundamental property considered in time series methods. The mean
is the first moment measured about the origin. It is also the average of all obser-
vations on a random variable. For a discrete time series Xi, i = 1, 2, …, N, the
sample mean X
(or population mean μ) is calculated as:

XN

¼1
X Xi ð3:1Þ
N i¼1

Although the mean conveys certain information, it does not completely char-
acterize a time series.

3.4.2 Variance

The variance is another basic property considered in time series methods. It is the
second moment about the mean, and is an indicator of the closeness of the values of
a time series to its mean. For a discrete time series Xi, i = 1, 2, …, N, the sample
variance s2 (or population variance σ2) is given as:

1 X N

Þ2
s2 ¼ ðXi  X ð3:2Þ
N  1 i¼1

Although variance is used in all aspects of statistical analysis, its use as a

descriptor is limited because of its units. Specifically, the units of the variance are
not the same as those of the time series. The square root of variance is the standard
deviation, s.
3.4 Relevant Statistical Characteristics and Estimators 69

3.4.3 Coefficient of Variation

The coefficient of variation (CV) or the relative standard deviation (RSD) of a time
series is defined as the ratio of the standard deviation to the mean of the time series.
It represents the extent of variability in relation to the mean of the time series. It is
written as:
s
CV ¼
ð3:3Þ
X

The coefficient of variation is also particularly useful when comparing the

variability of different time series. For instance, considering two time series, the
time series with a higher level of variability will have a greater CV value than the
other. This is a significant advantage in using CV, since the mean and the variance
cannot be used to actually compare the variability of two time series.

3.4.4 Skewness Coefficient

The skew is the third moment measured about the mean. For a discrete time series
Xi, i = 1, 2, …, N, the sample skewness coefficient g (or population skewness
coefficient γ) is given as:
P

Þ3
N Ni¼1 ðXi  X
g¼ ð3:4Þ
ðN  1ÞðN  2Þs3

The skew is a measure of symmetry. For a symmetric distribution, the skew will
be zero. For a nonsymmetric distribution, the skew will be positive or negative
depending on the location of the tail of the distribution. If the more extreme tail of
the distribution is to the right of the mean, the skew is positive. If the more extreme
tail of the distribution is to the left of the mean, the skew is negative.

3.4.5 Autocorrelation Function

The autocorrelation in a time series represents correlations between the values. It is

determined by the autocorrelation function, which is a normalized measure of the
linear correlation among successive values in a time series. Autocorrelation func-
tion is useful to determine the degree of dependence present in the values. For a
discrete time series Xi, i = 1, 2, …, N, and for a given lag time τ, the sample
70 3 Stochastic Time Series Methods

autocorrelation function rτ (or the population autocorrelation function ρτ) is given

by:
PNs PNs PNs
Xi Xi þ s  Ns 1
i¼1 Xi þ s i¼1 Xi
rs ¼ h 
i¼1
 i h PNs 2 i1=2
PNs 2 P Ns 2 1=2 P Ns 2
i¼1 Xi  1
Ns i¼1 X i i¼1 Xiþs  1
Ns i¼1 Xiþs

ð3:5Þ

In general, for a periodic process, the autocorrelation function is also periodic,

indicating the strong relation between values that repeat over and over again. For a
purely random process, the autocorrelation function fluctuates randomly about zero,
indicating that the process at any certain instance has no ‘memory’ of the past at all.
Other stochastic processes generally have decaying autocorrelations, but the rate of
decay depends on the properties of the process.
The autocorrelation function has been widely used for analysis of hydrologic
time series, including for identification of relevant antecedent conditions and
temporal persistence governing the process (e.g. Matalas 1963a; Salas 1993); see
also Salas et al. (1995). However, autocorrelations are not characteristic enough to
distinguish random from deterministic chaotic signals; see Chap. 6 for additional
details.

3.4.6 Power Spectrum

The spectral analysis is most useful in isolating periodicities in a time series, which
are often best delineated by analyzing the data in the frequency domain. The power
spectrum is a widely used tool for studying the oscillations of a time series. It is
defined as the square of the coefficients in a Fourier series representation of the time
series. It shows the variance of the function at different frequencies. For a discrete
time series Xi, i = 1, 2, …, N, if the power spectrum P(f) obeys a power law form

Pðf Þ / f b ð3:6Þ

where f is the frequency β is the spectral exponent, this is an indication of the

absence of characteristic timescale in the range of the power law. In such a case, a
fractal or scale-invariant behavior may be assumed to hold (e.g. Fraedrich and
Larnder 1993).
For a purely random process, the power spectrum oscillates randomly about a
constant value, indicating that no frequency explains any more of the variance of
the sequence than any other frequency. Periodic or quasi-periodic signals show
sharp spectral lines; measurement noise adds a continuous floor to the spectrum.
Thus, in the spectrum, signal and noise are readily distinguished.
3.4 Relevant Statistical Characteristics and Estimators 71

Spectral analysis, and especially the power spectrum, has been widely applied to
study hydrologic time series, including for identification of temporal persistence
and scale-invariant behavior (e.g. Roesner and Yevjevich 1966; Quimpo 1967;
Fraedrich and Larnder 1993; Olsson et al. 1993; Tessier et al. 1996; Menabde et al.
1997; Pelletier and Turcotte 1997; Sivakumar 2000; Mathevet et al. 2004).
However, power spectrum has limited ability in distinguishing between noise and
chaotic signals, since the latter can also have sharp spectral lines but even in the
absence of noise can have broadband spectrum; see Chap. 6 for additional details.
Basically, both the autocorrelation function and the spectral density function
contain the same information (since the spectral density is defined as the Fourier
transform of the autocorrelation function). The difference is that this information is
presented in the time (or space) domain by the autocorrelation function and in the
frequency domain by the spectral density function.

3.5 Parametric Methods

In parametric methods, the structure of the model is specified a priori, and the
number and nature of the parameters are generally fixed in advance. A large number
of parametric methods have been developed in the literature. These methods include
autoregressive (AR), moving average (MA), autoregressive moving average
(ARMA), autoregressive integrated moving average (ARIMA), gamma autorgres-
sive (GAR), periodic counterparts of AR, ARMA, and GAR, disaggregation,
fractional Gaussian noise (FGN), point process, Markov chain process, scaling, and
others. Extensive details of these models are available in Box and Jenkins (1970),
Brillinger (1975), Chatfield (1989), and Brockwell and Davis (1991), among others.
These methods have been extensively applied in hydrology for studying a wide
variety of time series and associated problems; see Clarke (1994), Haan (1994),
Hipel and McLeod (1994), and Salas et al. (1995) for details. Here, a brief account
of some of these methods and their applications in hydrology is presented.

3.5.1 Autoregressive (AR) Models

In an autoregressive (AR) model, the present outcome is considered to be a linear

combination of the signal in the past (with a finite memory) plus additive noise. For
a time series, Xi, an autoregressive model of order p, i.e. AR(p) model or simply the
AR model, is given by (e.g. Box and Jenkins 1970; Salas et al. 1995):
 
Xi ¼ l þ /1 ðXi1  lÞ þ /2 ðXi2  lÞ þ    þ /p Xip  l þ ei ð3:7Þ
72 3 Stochastic Time Series Methods

where μ is the mean value of the series, ϕ is the regression coefficient, and εi is the
error or noise, which is usually assumed to be an uncorrelated normal random
variable with mean zero and variance r2e (i.e. white Gaussian noise). The error εi is
also uncorrelated with Xi–1, Xi–2, …, Xi–p. Since εi is normally distributed, Xi is also
normally distributed. In order to determine the order p of the autoregression
required to describe the persistence adequately, it is necessary to estimate p + 2
parameters: ϕ1, ϕ2, …, ϕp, μ, and r2e . Several methods have been proposed in the
literature for an efficient estimation of these parameters; see Jenkins and Watts
(1968) and Kendall and Stuart (1968) for some early studies.
The mean, variance, and autocorrelation function of the AR(p) process are (e.g.
Box and Jenkins 1970; Salas et al. 1995)

EðXÞ ¼ l ð3:8aÞ

r2
VarðXÞ ¼ r2 ¼ Ppe ð3:8bÞ
1 j¼1 /j qj

qs ¼ /1 qs1 þ /1 qs2 þ    þ /p qsp ð3:8cÞ

respectively. Equation (3.8c) is also known as the Yule–Walker equation. Given

the model parameters, these three equations are useful for determining the prop-
erties of a model. Similarly, given a set of observations X1, X2, …, XN, they are
useful for estimating the parameters of the model.
The simplest form of the AR model when p = 1, i.e. AR(1) model, is given by:

Xi ¼ l þ /1 ðXi1  lÞ þ ei ð3:9Þ

Equation (3.9) is popularly known as the first-order Markov model. Such a model
states that the value of X in one time period is dependent only on the value of X in
the preceding time period plus a random component (with the random component
independent of X). The AR(1) model has three parameters to be estimated: μ, ϕ1,
and r2e . For this model, Eqs. (3.8b) and (3.8c) become

r2e
r2 ¼ ð3:10aÞ
1  /21

qs ¼ /1 qs1 ¼ /s1 ð3:10bÞ

Autoregressive models have been widely used in hydrology. In particular, the

first-order Markov model and other low-order AR models have been used for
modeling annual hydrologic time series as well as seasonal and daily time series
after standardization; see, for example, Thomas and Fiering (1962), Yevjevich
(1964, 1975), Matalas (1967), and Quimpo (1967) for some early studies.
3.5 Parametric Methods 73

3.5.2 Moving Average (MA) Models

The moving average (MA) is used to smooth various types of time series. The
moving average process used in the stochastic generation of hydrologic time series
is somewhat different. In this, the moving average process describes the deviations
of a sequence of events from their mean value. For a time series, Xi, a moving
average model of order q, i.e. MA(q) model or simply the MA model, is given by
(e.g. Box and Jenkins 1970):

Xi ¼ l þ ei  h1 ei1  h2 ei2      hq eiq ð3:11Þ

The simplest form of the MA model when q = 1, i.e. MA(1) model, is given by:

Xi ¼ l þ ei  h1 ei1 ð3:12Þ

For this model,

 
r2 ¼ r2e 1  h21 ð3:13aÞ

h1
qs ¼ ð3:13bÞ
1  h21

The moving average model has been applied in many early studies on hydro-
logic time series analysis, especially to study annual series. For instance, Matalas
(1963b) used the MA model to relate the effective annual precipitation and the
annual runoff. Yevjevich (1963) used the MA model to relate the mean annual
runoff to the annual effective precipitation.
Although useful, the MA model has not been particularly effective in the
analysis of many hydrologic time series when applied independently. On the other
hand, the model has been found to be very useful when combined with some other
models, such as the autoregressive model, as described below.

3.5.3 Autoregressive Moving Average (ARMA) Models

The autoregressive moving average (ARMA) models combine any direct auto-
correlation properties of a time series with the smoothing effects of an updated
running mean through the series. The ARMA(p,q) model or simply the ARMA
model is defined as (e.g. Box and Jenkins 1970; Salas et al. 1995):
 
Xi ¼ l þ /1 ðXi1  lÞ þ    þ /p Xip  l þ ei  h1 ei1      hq eiq ð3:14Þ
74 3 Stochastic Time Series Methods

with p autoregressive parameters ϕ1, ϕ2, …, ϕp, and q moving average parameters
θ1, θ2, …, θq. As above, the noise εi is an uncorrelated normal process with mean
zero and variance σ2ε and is also uncorrelated with Xi–1, Xi–2, …, Xi–p. An ARMA
(p,0) model is the same as an AR(p) model and an ARMA(0,q) model is the same
as the MA(q) model.
One of the merits of the ARMA process is that, in general, it is possible to fit a
model with a small number of parameters, i.e. p + q. This number is generally
smaller than the number of parameters that would be necessary using either an AR
model or an MA model. This principle is called the parsimony of parameters.
A simple form of the ARMA model with p = 1 and q = 1, i.e. the ARMA(1,1)
model, is given by:

Xi ¼ l þ /1 ðXi1  lÞ þ ei  h1 ei1 ð3:15Þ

with −1 < ϕ1 < 1 and −1 < θ1 < 1. The variance and lag-1 autocorrelation coeffi-
cient of the ARMA(1,1) model are:

1  2/1 h1 þ h21 2
r2 ¼ re ð3:16aÞ
1  /21

ð1  /1 h1 Þð/1  h1 Þ
q1 ¼ ð3:16bÞ
1  2/1 h1 þ h21

respectively. Furthermore, the autocorrelation function is:

qs ¼ /1 qs1 ¼ q1 /s1
1 s[1 ð3:16cÞ

Comparing Eqs. (3.16c) and (3.10b), one may observe that ρτ of the AR(1)
process is less flexible than that of the ARMA(1,1) process, since the former
depends on the sole parameter ϕ1, while the latter depends on ϕ1 and θ1. In general,
AR processes are short memory processes, while ARMA processes are
long-memory processes (Salas et al. 1979; Salas 1993).
The ARMA(1,1) model and other low-order ARMA models are very useful in
hydrology. They have found widespread applications in hydrology; see Carlson
et al. (1970), McKerchar and Delleur (1974), Moss and Bryson (1974), Tao and
Delleur (1976), Hipel et al. (1977), Lettenmaier and Burges (1977), McLeod et al.
(1977), Delleur and Kavvas (1978), Hipel and McLeod (1978a), Cooper and Wood
(1982), Salas and Obeysekera (1982), and Stedinger and Taylor (1982) for some
early applications. The applications have been on annual hydrologic time series,
seasonal time series after seasonal standardization, daily time series either after
seasonal standardization or by separating the year into several seasons and applying
different models to the daily series in each season, and many others.
3.5 Parametric Methods 75

3.5.4 Gamma Autoregressive (GAR) Models

Since the AR, MA, and ARMA models are based on the assumption that the error
εi, and, hence, the time series Xi are normally distributed, they cannot be directly
applied to skewed hydrologic time series. Their application to skewed time series
requires transformation of the time series into Normal processes. The gamma
autoregressive (GAR) model does not require such a transformation and, thus,
offers a direct modeling approach for skewed time series. The GAR model is based
on the assumption that the underlying series have a gamma marginal distribution.
The GAR model is defined as (Lawrance and Lewis 1981):

Xi ¼ /Xi1 þ ei ð3:17Þ

where ϕ is the autoregressive coefficient, εi is a random component, and Xi has a

three-parameter gamma marginal distribution given by the function:

ab ðx  kÞb1 exp½aðx  kÞ

fx ðxÞ ¼ ð3:18Þ
CðbÞ

where λ, α, and β are the location, scale, and shape parameters, respectively.
There are two ways to obtain the variable εi. For integer values of β, ε is given by
(Gaver and Lewis 1980):

kð1  /Þ X
b
e¼ þ gj ð3:19Þ
b j¼1

where ηj = 0, with probability ϕ; ηj = exp(α), with probability (1 − ϕ); and exp(α)

is an exponentially distributed random variable with expected value 1/α. This
approach is valid for skewness coefficient less than or equal to 2.0. For non-integer
values of β, ε is given by (Lawrance 1982):

e ¼ kð1  /Þ þ g ð3:20Þ

where

g ¼ 1 if M ¼ 0 ð3:21aÞ

X
M
g¼ Ej /Uj if M [ 0 ð3:21bÞ
j¼1
76 3 Stochastic Time Series Methods

where M is an integer random variable with Poisson distribution of mean value

−βln(ϕ). The set U1,…, UM are independent and identically distributed (iid) random
variables with uniform distribution (0,1) and the set E1,…, EM are iid random
variables exponentially distributed with mean 1/α (Salas 1993).
The GAR model has been applied in many hydrologic studies, including for
modeling annual streamflow and rainfall disaggregation; see Fernandez and Salas
(1990), Cigizoglu and Bayazit (1998), and Koutsoyiannis and Onof (2001), among
others.

3.5.5 Periodic Models: PAR, PARMA, and PGAR Models

To incorporate periodicity in processes (e.g. seasonality in rainfall and streamflow),

a number of periodic models have been suggested in the literature; see Box and
Jenkins (1970), Brockwell and Davis (1991), and Salas et al. (1995) for extensive
details. Of particular significance are the periodic counterparts of AR, ARMA, and
GAR models. In what follows, periodic AR (PAR), periodic ARMA (PARMA) and
multiplicative PARMA, and periodic GAR (PGAR) models are described, as these
models have found widespread applications in hydrology; see Loucks et al. (1981),
Salas (1993), Hipel and McLeod (1994), and Salas et al. (1995) for some com-
prehensive accounts.

3.5.5.1 Periodic AR (PAR) Model

Let us consider a periodic time series represented by Xy,s, y = 1, 2, …, N; s = 1, 2,

…, ω, where y is the year and s is the season, N is the number of years of record,
and ω is the number of seasons in a year. Here, s can represent a day, week, month,
or season. If s represents a month, then ω = 12. The periodic autoregressive model
of order p, i.e. PAR(p) model, for such a time series is defined as:

X
p
 
Xy;s ¼ ls þ /j;s Xy;sj  lsj þ ey;s ð3:22Þ
j¼1

where εy,s is an uncorrelated normal variable with mean zero and variance σ2s (ε), and it
is also uncorrelated with Xy,s−1, Xy,s−2, …, Xy,s−p. The model parameters are μs, ϕ1,s,
ϕ2,s, …, ϕp,s and σ2s (ε) for s = 1, 2, …, ω. In Eq. (3.22), if s – j ≤ 0, then Xy,s−j
becomes Xy−1,ω+s−j and μs−j becomes μω+s−j. The simplest PAR(p) model, i.e. PAR(1)
model, can be written as:
 
Xy;s ¼ ls þ /1;s Xy;s1  ls1 þ ey;s ð3:23Þ
3.5 Parametric Methods 77

The low-order PAR models have been widely used in hydrology, especially for
monthly (and seasonal) rainfall and streamflow simulations; see Hannan (1955),
Thomas and Fiering (1962), Fiering and Jackson (1971), Delleur et al. (1976), Salas
and Abdelmohsen (1993), and Shahjahan Mondal and Wasimi (2006), among
others.

3.5.5.2 Periodic ARMA (PARMA) Model

In a similar manner, the periodic autoregressive moving average model, i.e.

PARMA(p,q) model, for the above time series Xy,s is defined, by extending the
PAR(p) model, as:

X
p
  X
q
Xy;s ¼ ls þ /j;s Xy;sj  lsj þ ey;s  hj;s ey;sj ð3:24Þ
j¼1 j¼1

The model parameters are μs, ϕ1,s, ϕ2,s, …, ϕp,s, θ1,s, θ2,s, …, θp,s, and σ2s (ε) for
s = 1, 2, …, ω. When q = 0, the model becomes the well-known PARMA(p,0) or
PAR(q) model. The simplest PARMA(p,q) model, i.e. PARMA(1,1) model, can be
written as:
 
Xy;s ¼ ls þ /1;s Xy;s1  ls1 þ ey;s  h1;s ey;s1 ð3:25Þ

Low-order PARMA models, such as PARMA(1,0) and PARMA(1,1), have been

widely used for modeling seasonal hydrologic processes, especially for simulating
monthly and weekly flows; see Delleur and Kavvas (1978), Hirsch (1979), Vecchia
(1985), Ula (1990), Salas and Obeysekera (1992), Bartolini and Salas (1993), Salas
(1993), Salas and Abdelmohsen (1993), Rasmussen et al. (1996), and Tesfaye et al.
(2006), among others. In such studies, the method of moments is typically used for
estimation of the model parameters.
Physically-based or conceptual arguments of the hydrologic cycle of a watershed
or river basin justify the applicability of the PARMA models. For instance, Salas
and Obeysekera (1992) showed that assuming that the precipitation input is an
uncorrelated periodic-stochastic process and under some linear reservoir consid-
erations for the groundwater storage, the stochastic model for seasonal streamflow
becomes a PARMA(1,1) process. Some studies have also suggested a constant
parameter ARMA(2,2) model with periodic independent residuals (e.g. Claps et al.
1993; Murrone et al. 1997).

3.5.5.3 Multiplicative PARMA Model

It is obvious that preservation of both seasonal and annual statistics should be a

desirable property of periodic stochastic models. However, such dual preservation
78 3 Stochastic Time Series Methods

of statistics is also difficult to achieve with low-order PAR models, such as the
PAR(1) and PAR(2). While PARMA models offer the possibility of preserving both
seasonal and annual statistics, due to their more flexible correlation structure, there
are some concerns that such models have too many parameters (although the
number may be reduced by keeping some of the parameters constant). To overcome
these problems, the family of multiplicative models was proposed as an alternative
(Box and Jenkins 1970).
Multiplicative models have the characteristic of linking the variable Xy,s with Xy,s−1
and Xy−1,s. However, without periodic parameters, such models often cannot produce
the seasonality in the covariance structure of the process, as was shown by McKerchar
and Delleur (1974) in their study on simulation and forecasting of monthly stream-
flows. A model, with periodic parameters, that can overcome these limitations is the
multiplicative PARMA model. The multiplicative PARMA(1,1) × (1,1)ω model is
given by:
     
Xy;s ¼ ls þ U1;s Xy1;s  ls þ /1;s Xy;s1  ls1  U1;s /1;s Xy1;s1  ls1
þ ey;s  H1;s ey1;s  h1;s ey;s1 þ H1;s h1;s ey1;s1
ð3:26Þ

where U1;s , H1;s , h1;s , and σs(ε) are the model parameters.

3.5.5.4 Periodic GAR (PGAR) Model

The above PARMA and multiplicative PARMA models for modeling periodic time
series require transformation of the time series into Normal. An alternative that can
overcome this problem is the periodic GAR model of order 1, i.e. PGAR(1) model.
This model has periodic correlation structure and periodic gamma marginal dis-
tribution (e.g. Fernandez and Salas 1986). Let us consider that Xy,s is a periodic
correlated variable with gamma marginal distribution with location λs, scale αs, and
shape βs parameters varying with s, and s = 1, 2, …, ω. Then, the variable Zy,s =
Xy,s − λs is a two-parameter gamma that can be represented by the model
   d
Xy;s ¼ ks þ /s Xy;s1  ks1 þ Xy;s1  ks1 s Wy;s ð3:27Þ

where /s is the periodic autoregressive coefficient, δs is the periodic autoregressive

exponent, and Wy,s is the noise process. This model has a periodic correlation
structure equivalent to that of the PAR(1) process. An early application of this
model in hydrology was made by Fernandez and Salas (1986) for modeling the
weekly streamflows in several rivers in the United States. The model favorably
compared with respect to the Normal-based models (e.g. the PAR model after
3.5 Parametric Methods 79

logarithmic transformation) in reproducing the basic statistics usually considered

for streamflow simulation.

3.5.6 Extension of AR, ARMA, PAR, and PARMA Models

to Multiple Variables

3.5.6.1 Multivariate AR and Multivariate ARMA Models

Let us consider multiple time series X(1)i , Xi , …, Xi , with n representing the

(2) (n)

number of variables or sites. This multiple time series can be represented as a

column vector Xi, with elements X(1)i , Xi , …, Xi . With this, the simplest multi-
(2) (n)

variate AR model, i.e. multivariate AR(1) model, can be written as (Matalas 1967):

X i ¼ l þ A1 ðXi1  lÞ þ Bei ð3:28Þ

where μ is a column parameter vector with elements μ(1) i , μi , …, μi ; and A1 and

(2) (n)

B are n × n parameter matrices. The noise term εi is also a column vector of noises
ε(1)
i , εi , …, εi , each with zero mean such that E(εiεi ) = I, where T denotes the
(2) (n) T

transpose of the matrix and I is the identity matrix, and E(εiεTi−τ) = 0 for τ ≠ 0. In
addition, εi is uncorrelated with Xi−1 and εi is also normally distributed.
In a similar manner, the simplest multivariate ARMA model, i.e. multivariate
ARMA(1,1) model, is given by:

Xi ¼ l þ A1 ðXi1  lÞ þ Bei  C1 ei1 ð3:29Þ

where C1 is an additional n × n parameter matrix.

The multivariate AR and multivariate ARMA models have found a number of
applications in hydrology; see Pegram and James (1972), Ledolter (1978), Cooper
and Wood (1982), Stedinger et al. (1985), and Chaleeraktrakoon (1999, 2009),
among others; see also Salas et al. (1995) for a comprehensive account. A particular
limitation in using the full multivariate AR and multivariate ARMA models is that
they often lead to complex parameter estimation. To overcome this issue, simpli-
fications have been suggested. One such simplification is the assumption of a
diagonal matrix for A1 in the multivariate AR model (Eq. 3.28) (e.g. Matalas 1967)
and for A1 and C1 in the multivariate ARMA model (Eq. 3.29) (e.g. Salas et al.
1995). These simplified models are called contemporaneous models, to imply that
only the dependence of concurrent values of the X’s are considered important. The
diagonalization of the parameter matrices allows model decoupling into component
univariate models, so that the model parameters do not have to be estimated jointly
and that univariate modeling procedures can be employed. The contemporaneous
AR model is termed as the CAR model and the contemporaneous ARMA model is
popularly known as the CARMA model.
80 3 Stochastic Time Series Methods

3.5.6.2 Multivariate PAR and Multivariate PARMA Models

The multivariate AR and multivariate ARMA models can be modified for periodic
series, similar to the modifications presented earlier for the univariate periodic
y,s , Xy,s , …, Xy,s , where y is
series. Let us consider a periodic multiple time series X(1) (2) (n)

the year, s is the season, and n is the number of variables or sites. These time series
can be represented as a column vector Xy,s, with elements X(1) y,s , Xy,s , …, Xy,s . With
(2) (n)

this, the simplest multivariate periodic AR model, i.e. multivariate PAR(1) model,
can be written as:
 
X y;s ¼ ls þ As Xy;s1  ls þ Bs ey;s ð3:30Þ

where μ is a column parameter vector with elements μ(1) s , μs , …, μs ; and As and

(2) (n)

Bs are n × n parameter matrices. All these three parameters are periodic. The noise
term εy,s is also a column vector of noises ε(1) y,s , εy,s , …, εy,s , each with zero mean
(2) (n)

such that E(εy,sεy,s) = I, and E(εy,sεy,s-τ) = 0 for τ ≠ 0. In addition, εy,s is uncorrelated

T T

with Xy,s–1 and εy,s is also normally distributed.

In a similar manner, the simplest multivariate periodic ARMA model, i.e.
multivariate PARMA(1,1) model, is given by:
 
X y;s ¼ ls þ As X y;s1  ls þ Bs ey;s  Cs ey;s1 ð3:31Þ

where Cs is an additional n × n periodic parameter matrix. Simplifications of these

models can be made through diagonalization. These simplified models are called
contemporaneous periodic AR model and contemporaneous periodic ARMA
model, respectively, for PAR and PARMA models.
A number of studies have used the multivariate PAR and multivariate PARMA
and their contemporaneous versions for modeling hydrologic time series; see
Vecchia (1985), Bartolini et al. (1988), Ula (1990), Salas and Abdelmohsen (1993),
and Rasmussen et al. (1996), among others.

3.5.7 Disaggregation Models

Many hydrologic design and operational problems often require data at much finer
scales than that are commonly available through measurements. For instance,
studies on floods and design of urban drainage structures require rainfall data at
hourly scale or at even finer scales. However, rainfall data are widely available only
at the daily scale. Therefore, it becomes necessary to disaggregate the available
daily rainfall data to hourly or finer-scale data. For streamflow, the problem may be
to obtain monthly flows from annual flows, since simulations of annual flows are
much more accurate than those of monthly flows. These data issues are applicable
3.5 Parametric Methods 81

not only to time but also to space and, consequently, to space-time. In view of these,
development and applications of disaggregation models have been an important
part of stochastic hydrology. Such disaggregation models can be broadly catego-
rized into temporal disaggregation models, spatial disaggregation models, and
space-time disaggregation models. Here, a brief account of such models is pre-
sented, with a description of the Valencia-Schaake model (Valencia and Schaake
1973) for disaggregation of streamflows from the annual scale to the seasonal scale
(e.g. monthly), as this model has been widely used in hydrology, especially in
streamflow studies.
The Valencia-Schaake model to disaggregate streamflows from the annual scale
to the seasonal scale (with number of seasons s) at n sites is written as:

Y ¼ AX þ Be ð3:32Þ

where Y is an ns vector of disaggregated seasonal values (monthly streamflows),

X is an n vector of aggregate annual values (annual streamflows), A and B are
ns × ns parameter matrices, and ε is an ns vector of independent standard normal
variables. The matrix A is estimated to reproduce the correlation between aggregate
and disaggregate values, whereas the matrix B is estimated to produce the corre-
lation between individual disaggregate components. Parameter estimation, based on
the method of moments, generally leads to the preservation of the first- and
second-order moments at all levels of aggregation. In this model, the derivation of
the monthly streamflow values is accomplished in two or more steps. First, the
aggregate annual values are modeled so as to reproduce the desired annual statistics
(e.g. based on the ARMA(1,1) model). Then, synthetic annual values are generated
and subsequently disaggregated into the seasonal values by means of the model
parameters in Eq. (3.32).
The Valencia-Schaake model generally performs well for disaggregation, as the
variance-covariance properties of the seasonal data are preserved and the generated
seasonal values also add up to the annual values. However, it also suffers from the
following shortcomings, among others: (1) the model does not preserve the
covariances of the first season of a year and any preceding season; (2) the model is
compatible mainly with data that exhibit Gaussian distributions, but some hydro-
logic data (including daily and sub-daily rainfall data as well as monthly streamflow
data) are seldom normally distributed. Therefore, it becomes necessary to apply
some kind of normalizing transformations to the data before its application.
However, it is often difficult to find a general normalizing transformation and retain
the statistical properties of the process; (3) the model involves an excessive number
of parameters; and (4) the linear nature of the model limits it from representing any
nonlinearity in the dependence structure between variables, except through the
normalizing transformation used.
Some of these limitations have been addressed by certain modifications to the
model. For instance, Mejia and Rouselle (1976) presented a modification to
82 3 Stochastic Time Series Methods

preserve the covariances of the first season of a year and any preceding season.
Their model is given by:

Y ¼ AX þ Be þ CZ ð3:33Þ

where C is an additional parameter matrix and Z is a vector of seasonal values from

the previous year (usually only the last season of the previous year). The issues
regarding summability (i.e. the requirement that disaggregate variables should add
up to the aggregate quantity), Gaussian assumption, and number of parameters have
been addressed by many studies through other modifications (e.g. Tao and Delleur,
1976; Hoshi and Burges 1979; Lane 1979; Todini 1980; Stedinger and Vogel 1984;
Stedinger et al. 1985; Grygier and Stedinger 1988; Koutsoyiannis 1992; Santos and
Salas 1992; Salas 1993). These limitations also led to the proposal of new disag-
gregation models that do not necessarily exhibit the generality of the
Valencia-Schaake model. Such models are based on non-dimensionalized Markov
process (e.g. Woolhiser and Osborne 1985; Hershenhorn and Woolhiser 1987),
dynamic step-wise disaggregation (e.g. Koutsoyiannis and Xanthopoulos 1990;
Koutsoyiannis 2001), Bartlett-Lewis rectangular pulses (e.g. Bo et al. 1994;
Glasbey et al. 1995), Poisson cluster process (e.g. Connolly et al. 1998;
Koutsoyiannis and Onof 2001), scaling (Perica and Foufoula-Georgiou 1996;
Olsson 1998), and more general forms (e.g. Koutsoyiannis 2000). Furthermore, to
overcome the issues relating to a priori definition of the model structure in these
parametric models, nonparametric and semi-parametric approaches for disaggre-
gation have also been proposed (e.g. Tarboton et al. 1998; Nagesh Kumar et al.
2000; Srinivas and Srinivasan 2005; Prairie et al. 2007; Nowak et al. 2010); see also
Sect. 3.6 for a discussion of nonparametric approaches.

3.5.8 Markov Chain Models

A Markov chain is a discrete-time stochastic process that undergoes transitions from

one state to another on a state space. Let us consider a discrete-time process X(i)
developing through time i. Let us denote the values this process takes as Xi, i = 1, 2, …,
N. With this, given the entire history of the process, the probability of the process
being equal to Xi at time i can be written as:

P½XðiÞ ¼ Xi jXð1Þ ¼ X1 ; Xð2Þ ¼ X2 ; . . .; Xði  1Þ ¼ Xi1  ð3:34Þ

Simplification of Eq. (3.34) as:

P½XðiÞ ¼ Xi jXði  1Þ ¼ Xi1  ð3:35Þ

3.5 Parametric Methods 83

indicates that the outcome of the process at time i can be defined by using only the
outcome at time i − 1. A process with this property is known as the first-order
Markov chain or a simple Markov chain.
Let us consider that there is a total of u number of states of the process and that
any given state is denoted by l. With this, instead of using Xi, one can use X(i) to
represent the state of the process, such as X(i) = l, l = 1, 2, …, u, which means that
X(i) is at state l. For instance, for rainfall modeling, a total of two states (i.e. u = 2)
is normally considered: l = 1 for a dry state (i.e. no rain) and l = 2 for a wet state.
The first-order Markov chain is defined by its transition probability matrix P(i).
The matrix P(i) is a square matrix with elements pkl(i) and is given by:

pkl ðiÞ ¼ P½XðiÞ ¼ ljXði  1Þ ¼ k ð3:36Þ

for all k, l pairs. Therefore, one can write that

X
u
pkl ðiÞ ¼ 1 k ¼ 1; . . .; u ð3:37Þ
l¼1

If the transition probability matrix P(i) does not depend on time i, then the
Markov chain is a homogeneous chain or a stationary chain. In such a case, the
transition probability matrix is denoted as P and the elements are denoted as pkl.
The probabilities that are often useful in hydrology, especially in rainfall, are the r-
ðrÞ
step transition probability pkl (assuming that the chain changes from state k to state
l in r steps), the marginal distribution ql(i) given the distribution ql(1) (i.e. for the
initial state), and the steady-state probability vector q*; see Parzen (1962) for
details.
The Markov chain models have been widely used in hydrology, including for
rainfall (e.g. Gabriel and Neumann 1962; Haan et al. 1976; Buishand 1977; Chin
1977; Katz 1977; Roldan and Woolhiser 1982; Chang et al. 1984; Mimikou 1984;
Foufoula-Georgiou and Lettenmaier 1987; Bardossy and Plate 1991; Zucchini and
Guttorp 1991; Katz and Parlange 1995; Rajagopalan et al. 1996; Wilks 1998;
Hughes et al. 1999; Thyer and Kuczera 2000; Robertson et al. 2004; Lennartsson
et al. 2008), streamflow (e.g. Yevjevich 1972; Şen 1976, 1990; Chung and Salas
2000; Cancelliere and Salas 2004, 2010; Bayazit and Önöz 2005; Akyuz et al.
2012; Sharma and Panu 2014), and water storage (e.g. Moran 1954; Lloyd 1963;
Gani 1969), among others. In recent years, use of Markov chains for parameter
estimation and uncertainty in hydrologic models has also been gaining considerable
attention; see Bates and Campbell (2001), Marshall et al. (2004), Vrugt et al.
(2008), and Vrugt (2016) for some accounts.
84 3 Stochastic Time Series Methods

3.5.9 Point Process Models

The theory of point processes (e.g. Cox 1958; Neyman and Scott 1958; Bartlett
1963) has been one of the earliest tools for modeling rainfall (e.g. Le Cam 1958).
A fundamental assumption in rainfall models based on point processes is that the
occurrence of rain storms is a Poisson process. Assuming that the storm arrivals are
governed by a Poisson process, the number of storms Ni in a time interval (1,i)
arriving at a location is Poisson-distributed with a parameter λi (storm arrival rate).

P½Ni ¼ n ¼ ðki Þn expðki Þ n ¼ 1; 2; . . . ð3:38Þ

If n storms arrived in the interval (1,i) at times i1, …, in, then the number of
storms in any time interval I (= i/n) is also Poisson-distributed with parameter λI. It
is further assumed that the rainfall amount (R) associated with a storm arrival is
white noise (e.g. gamma-distributed) and that the number of storms Ni and amount
R are assumed to be independent. Thus, rainfall amounts r1, …, rn correspond to
storms occurring at times i1, …, in. Such a rainfall generating process is called
Poisson white noise (PWN) and is a simple example of a point process.
Under this formulation, the cumulative rainfall in the interval (1,i) is given by:

X
Ni
Zi ¼ Rj ð3:39Þ
j¼1

The cumulative rainfall Zi is called a compound Poisson process.

In the PWN model, the rainfall is assumed to occur instantaneously, so the
storms have zero duration, which is also unrealistic. Modifications to the PWN
model consider rainfall as an occurrence with a random duration E and intensity Y,
called the Poisson rectangular pulse (PRP) model (Rodriguez-Iturbe et al. 1984).
Commonly, E and Y are assumed to be independent and exponentially distributed.
In this formulation, n storms may occur at times i1, …, in with associated durations
and intensities (y1, e1), …, (yn, en), and the storms may overlap, so that the
aggregated process Xt becomes autocorrelated. The PRP model is better concep-
tualized than the PWN model, but it is still limited when applied to rainfall data.
In view of the limitations of the simple point process models, such as the PWN
and PRP, several alternative models, often with greater complexity and sophisti-
cation, have also been suggested. These include models based on Cox processes,
renewal processes, cluster processes, and others; see Le Cam (1958), Buishand
(1977), Kavvas and Delleur (1981), Smith and Carr (1983), Rodriguez-Iturbe et al.
(1984, 1986, 1987), Ramirez and Bras (1985), Foufoula-Georgiou and Guttorp
(1986), and Foufoula-Georgiou and Lettenmaier (1987) for some early applications
of these concepts for modeling rainfall. Comprehensive reviews and comparisons of
different models are also available in the literature; see, for instance, Onof et al.
(2000) for some details. Among these, the cluster-based models have been widely
3.5 Parametric Methods 85

used, and in particular the Neyman-Scott model and the Bartlett-Lewis model.
A brief account of these two models is presented next.

3.5.9.1 Neyman-Scott Model

The Neyman-Scott model for rainfall (Rodriguez-Iturbe et al. 1987) was based on
the cluster point process model of Neyman and Scott (1958). The model assumes
that there exists a generating mechanism called the “storm origin” in any storm
event which may be passing fronts or some other criteria for convection storms
from which rain cells develop. The Neyman-Scott model is described by three
independent elementary stochastic processes: (1) a process that sets the origins of
the storms; (2) a process that determines the number of rain cells generated by each
storm; and (3) a process that defines the origin of the cells. The origins of the storms
(L) are governed by a Poisson process with rate parameter λ. At a point on the
ground, the storm is conceptualized as a random number of rain cells C, which are
Poisson or geometrically distributed. The cell origins and positions (B) are inde-
pendently separated from the storm origin by distances which are exponentially
distributed with parameter β. No cell origins are assumed to be located at the storm
origin. A rectangular pulse is associated independently with each cell origin with
random duration, E, and with random intensity, Y. The duration and intensity are
assumed to be exponentially distributed with parameters η and 1/μx, respectively,
and are independent of each other. With these, the basic Neyman-Scott model
consists of five variables (origin of storms—L, number of cells—C, position of cells
—B, duration of cells—E, and intensity of cells—Y) and five parameters (λ, 1/μc, β,
η, 1/μx). Two storms and cells may overlap, and the total rainfall intensity at any
point in time i, i.e. Xi, is given by the sum of the intensities of the individual cells
active at time i. If the rainfall cell is described by an instantaneous random rainfall
depth, the resulting rainfall process is known as Neyman-Scott white noise
(NSWN). On the other hand, if the rainfall cell is a rectangular pulse, the rainfall
process is known as Neyman-Scott rectangular pulse (NSRP).
Parameter estimation of Neyman-Scott models has been extensively studied in
the hydrologic literature using the method of moments and other approaches (e.g.
Kavvas and Delleur 1981; Obeysekera et al. 1987; Entekhabi et al. 1989; Islam
et al. 1990; Cowpertwait 1995). A major problem with parameter estimation in
Neyman-Scott models is that parameters estimated based on data for one level of
aggregation (e.g. hourly) may be significantly different from those estimated from
data for another level of aggregation (e.g. daily). Weighted moments estimates of
various timescales in a least squares fashion have been proposed as an alternative
(e.g. Entekhabi et al. 1989). Constraints may be set on the parameters based on the
physical understanding of the process that can improve parameter estimation
(Cowpertwait and O’Connell 1997), as shown in a space-time cluster model
(Waymire et al. 1984). However, the difficulty in estimating the parameters even
when using physical considerations persists.
86 3 Stochastic Time Series Methods

Since the introduction of the Neyman-Scott model for rainfall modeling,

numerous studies have applied the model and its variants for rainfall modeling and
for other subsequent analysis. Details of such applications can be seen in
Cowpertwait (1991), Puente et al. (1993), Cowpertwait et al. (1996a, b, 2002,
2007), Calenda and Napolitano (1999), Favre et al. (2002), Evin and Favre (2008),
Leonard et al. (2008), and Burton et al. (2010), among others.

3.5.9.2 Bartlett-Lewis Model

The Bartlett-Lewis model for rainfall was originally proposed by Rodriguez-Iturbe

et al. (1987). The model works in a somewhat similar manner to that of the
Neyman-Scott model. Similar to the Neyman-Scott model, the Bartlett-Lewis
model represents the arrival of rain storms as a Poisson process with rate λ, with
each storm generating a cluster of cell arrivals. However, the clustering mechanism
assumes that the time intervals between successive cells (rather than the temporal
distances of the cells from their storm origin) are independent and identically
distributed random variables. The intervals between successive cells are assumed to
be exponentially distributed, so that cell arrivals constitute a secondary Poisson
process of rate β. Each cell is associated with a rectangular pulse of rain, of random
duration, E, and with random intensity, Y. In the simplest version of the model,
these are both assumed to be exponentially distributed with parameters η and 1/μx,
respectively, and are independent of each other. The cell origin process terminates
after a time that is also exponentially distributed with rate γ. This basic version,
therefore, has five parameters in total: λ, β, η, 1/μx, and γ.
Since the original model of Rodriguez-Iturbe et al. (1987), several modifications
have been suggested to the Bartlett-Lewis rainfall model. One early modification
was proposed by Rodriguez-Iturbe et al. (1988), which involves randomization of
the cell duration parameter and related temporal storm characteristics to enable
variation between storms. This model, called the Bartlett-Lewis Random Parameter
(BLRPR) model, extends the basic model by allowing parameter η, that specifies
the duration of cells, to vary randomly between storms. This is achieved by
assuming that the η values for different storms are independent and
identically-distributed random variables from a gamma distribution with index
(shape) α and rate parameter (scale) ν. This way, the model is re-parameterized in
such a way that, rather than keeping β (the cell arrival rate) and γ (the storm
termination rate) constant for each storm, the ratio of both these parameters to η is
kept constant. This means that, for higher η (i.e. typically shorter cell durations),
there are correspondingly shorter storm durations and shorter cell inter-arrival
times. This model improves the simulation of dry spell lengths. Other modifications
to the BLRP model include the Bartlett-Lewis rectangular pulse gamma model
(Onof and Wheater 1994), hybrid-based model (Gyasi-Agyei and Willgoose 1997),
spatial-temporal model based on Bartlett-Lewis process (Northrop 1998), the
Bartlett-Lewis Instantaneous Pulse (BLIP) model (Copertwait et al. 2007), and the
3.5 Parametric Methods 87

random parameter Bartlett-Lewis Instantaneous Pulse (BLIPR) model (Kaczmarska

et al. 2014), among others.
The original Bartlett-Lewis model and its modified versions have been applied in
numerous rainfall studies. Extensive details of such applications, including reviews
and comparisons of different stochastic methods, are available in Islam et al. (1990),
Onof and Wheater (1993), Copertwait (1995, 1998, 2004, 2010), Khaliq and
Cunnane (1996), Verhoest et al. (1997, 2010), Cameron et al. (2000), Onof et al.
(2000), Koutsoyiannis and Onof (2001), Smithers et al. (2002), Marani and Zanetti
(2007), Copertwait et al. (2011), Pui et al. (2012), Pham et al. (2013), and
Kaczmarska et al. (2014), among others.

3.5.10 Other Models

In addition to the models discussed above, a number of other stochastic parametric

models also exist for time series analysis. Such models include those that are either
variants/extensions of the models discussed above or based on very different
concepts. Examples of such models are autoregressive integrated moving average
(ARIMA) models, fractional Gaussian noise (FGN) models, broken line
(BL) models, and scaling models. All these models have found important appli-
cations in hydrology; see Mandelbrot and Van Ness (1968), Mandelbrot and Wallis
(1969), Carlson et al. (1970), Mandelbrot (1972), Mejia et al. (1972, 1974), Delleur
and Kavvas (1978), Hipel and McLeod (1978b), Salas et al. (1982), Gupta and
Waymire (1990), Ahn and Salas (1997), Montanari et al. (1997), Venugopal et al.
(1999), Deidda (2000), and Seed et al. (2000) for some earlier applications.
Among these models, the scaling-based models have been gaining far more
attention over the past two decades or so. The scaling-based models are generally
based on self-similar or scale-invariant structure of the underlying process, and
originate from turbulence theory (e.g. Mandelbrot 1974; Meneveau and Sreenivasan
1987). Comprehensive accounts of the relevance of scaling-based concepts in
hydrology and their applications can be found in Mandelbrot (1983), Gupta et al.
(1986), Rodriguez-Iturbe and Rinaldo (1997), and Sposito (2008), among others.
Scaling theories are viewed in many different ways and, consequently, many dif-
ferent types of models have been developed and applied in hydrology. For instance,
these models can be categorized into canonical and microcanonical, bounded and
unbounded, simple-scaling and multi-scaling. Extensive details of such models and
their applications in hydrology can be found in Schertzer and Lovejoy (1987),
Lovejoy and Schertzer (1990), Gupta and Waymire (1993), Olsson et al. (1993),
Over and Gupta (1996), Carsteanu and Foufoula-Georgiou (1996), Gupta et al.
(1994, 1996), Perica and Foufoula-Georgiou (1996), Menabde et al. (1997),
Venugopal et al. (1999), Veneziano et al. (2000, 2006), Ferraris et al. (2003), Molnar
and Burlando (2005), Dodov and Foufoula-Georgiou (2005), Marani and Zanetti
(2007), Pui et al. (2012), and Markonis and Koutsoyiannis (2015), among others.
88 3 Stochastic Time Series Methods

A more recent comprehensive account of scaling-based methods and their applica-

tions in hydrology can be found in Veneziano and Langousis (2010).

3.5.11 Remarks

While the commonly used parametric models, such as the ones discussed above, are
indeed useful for modeling hydrologic time series, they have certain important
limitations. For instance: (1) since single linear model is often fit to the entire data,
the significance of local neighborhood is not given due consideration; (2) since the
structure of the model is specified a priori and the number and nature of the
parameters are generally fixed in advance, there is little flexibility; (3) the parameter
estimation procedure is often complicated, especially when the number of param-
eters is large, which is also often the case; and (4) the models are often not able to
capture several properties of hydrologic time series, including asymmetric and/or
multimodal conditional and marginal probability distributions, persistent large
amplitude variations at irregular time intervals, amplitude-frequency dependence,
apparent long memory, nonlinear dependence between Xi versus Xi−τ for lag τ, and
time irreversibility, among others; see Yakowitz (1973), Jackson (1975), Kendall
and Dracup (1991), Lall and Sharma (1996) for some additional details. These
limitations led to the development and applications of nonparametric methods in
hydrology; see Yakowitz (1979, 1985, 1993), Adamowski (1985), Karlsson and
Yakowitz (1987), Bardsley (1989), Kendall and Dracup (1991), Smith (1991),
Smith et al. (1992), Lall and Sharma (1996), and Lall et al. (1996) for some early
studies. The next section presents an overview of nonparametric methods and their
hydrologic applications.

3.6 Nonparametric Methods

The nonparametric methods make no prior assumptions on the model structure.

Instead, the model structure is determined from the data. Although nonparametric
methods may also involve parameters, the number and nature of the parameters are
not fixed in advance and, thus, are flexible. The nonparametric methods approxi-
mate the conditional and marginal distributions of a time series and simulate from
these. For instance, for a time series Xi, i = 1, 2, …, N, an order p model can be
expressed as a simulation from a conditional probability density function
 
  f Xi ; Xi1; Xi2; . . .; Xip
f Xi jXi1 ; Xi2 ; . . .; Xip ¼ R   ð3:40Þ
f Xi ; Xi1 ; Xi2 ; . . .; Xip dXi
3.6 Nonparametric Methods 89

This approach is different from the one adopted in parametric models, which
essentially simulate from a Gaussian conditional distribution because they assume
the data is normally distributed.
There exist many nonparametric methods in the literature; see Efron and
Tibishirani (1993), Higgins (2003), Sprent and Smeeton (2007), and Hollander
et al. (2013) for details. Among the nonparametric methods that have found
applications in hydrology are the nearest neighbor resampling, block bootstrap
resampling, kernel density estimator (KDE), k-nearest neighbor bootstrap resam-
pling (KNNR), k-nearest neighbor with local polynomial regression (LPK), non-
parametric order p simulation with long-term dependence (NPL), and hybrid
models. Details of these methods and applications in hydrology are available in
Yakowitz (1979), Lall and Sharma (1996), Vogel and Shallcross (1996), Sharma
et al. (1997), Tarboton et al. (1998), Rajagopalan and Lall (1999), Sharma and
O’Neill (2002), Srinivas and Srinivasan (2005, 2006), and Prairie et al. (2006),
among others. Some of these methods are briefly described next.

3.6.1 Bootstrap and Block Bootstrap

The bootstrap (Efron 1979; Efron and Tibishirani 1993) is perhaps the simplest
nonparametric technique for time series analysis. The bootstrap is a statistical
method that involves resampling the original time series (with replacement) to
estimate the distribution of a statistic (e.g. mean, variance, correlation). The clas-
sical idea is to resample the original time series (with replacement) to generate
B bootstrap samples, from which one can simulate B estimates of a given statistic,
leading to an empirical probability distribution of the statistic. Let us consider a
time series Xi, i = 1, 2, … N, denoted as x and the task is to estimate the empirical
probability distribution of a statistic ^hi . Each observation Xi is resampled (with
replacement) with an equal probability of 1/N. The sample x continues to be
resampled with replacement B times, until B bootstrap samples xi, i = 1, 2, …, B are
obtained. Each bootstrap sample xi yields a bootstrap estimate of the statistic θ
leading to the B bootstrap estimates ^hi ; i ¼ 1; 2; . . .; B.
By simply resampling (with replacement) from the original time series, the
bootstrap can be used as a nonparametric time series model. A particular challenge
in such a model, however, is to resample the time series to preserve the temporal
correlation ρ (and spatial covariance) structure of the original time series. This is
because the classic bootstrap assumes that the data are independent and identically
distributed and resamples from each prior data point with equal probability. One
way to address this problem is by resampling λ-year blocks, so that the resulting
sequence will be approximately independent and, thus, can preserve the serial
correlation structure of the time series. This kind of resampling is known as the
moving-blocks bootstrap (e.g. Künsch 1989; Efron and Tibshirani 1993). Here, a
90 3 Stochastic Time Series Methods

block length λ ≈ N/K, where N is the length of the time series and K is the number
of blocks to resample, is chosen, as opposed to a single observation, in the
bootstrap. Sampling blocks of length λ allows one to retain the original correlation ρ
among the observations within each block, and yet adjacent blocks are uncorrelated.
The basic idea is to choose a large enough block length λ so that observations more
than λ time units apart will be nearly independent.
Numerous studies have applied the bootstrapping, block bootstrapping, and
moving block bootstrapping, and related techniques in hydrology; see Labadie et al.
(1987), Zucchini and Adamson (1989), Kendall and Dracup (1991), Vogel and
Shallcross (1996), Ouarda et al. (1997), Kundzewicz and Robson (2004), Yue and
Pilon (2004), Noguchi et al. (2011), Önöz and Bayazit (2012), Sonali and Nagesh
Kumar (2013), and Hirsch et al. (2015), among others.

3.6.2 Kernel Density Estimate

Kernel density estimation entails a weighted moving average of the empirical

frequency distribution of the time series. Most nonparametric density estimators can
be expressed as kernel density estimators (Scott 1992).
For a univariate time series of observations Xi, i = 1, 2, …, N, the kernel
probability density estimator at any point X is written as:

X  
^f ðXÞ ¼
N
1 X  Xi
K ð3:41Þ
i¼1
Nki ki

where K(.) is a kernel function centered on the observation Xi that is usually taken
to be a symmetric, positive, probability density function with finite variance, and λ
is a bandwidth or “scale” parameter of the kernel centered at Xi. A fixed kernel
density estimator uses a constant bandwidth, λ, irrespective of the location of
X. Such a fixed estimate is formed by summing kernels with bandwidth λ centered
at each observation Xi, as given by:

X  
^f ðXÞ ¼
N
1 X  Xi
K ð3:42Þ
i¼1
Nk k

The kernel K(.) is a symmetric function centered on the observation Xi, that is
positive, integrates to unity, has first moment equal to zero and finite variance. This
is similar to the histogram construction, where individual observations contribute to
the density by placing a rectangular box (analogous to the kernel function) in the
prespecified bin the observation lies in (Sharma et al. 1997). The histogram is
sensitive to the position and size of each bin. Use of smooth kernel functions makes
the kernel density estimate in Eq. (3.41) smooth and continuous.
3.6 Nonparametric Methods 91

There are many possible kernel functions, including uniform, triangular, Normal,
Epanechnikov, Bisquare, and others; see Silverman (1986) and Scott (1992) for
details. However, the Gaussian kernel function is widely used. It is given by:
 
1 X2
KðXÞ ¼ exp  ð3:43Þ
ð2pÞ1=2 2

The univariate kernel density estimate in Eq. (3.42) can be easily extended to a
multivariate one with a dimension d. For instance, using a Gaussian kernel function,
the multivariate kernel probability density ^f ðXÞ of a d-dimensional variable set X is
estimated as:
!
XN
ðX  Xi ÞT S1 ðX  Xi Þ
^f ðXÞ ¼ 1 1
exp  ð3:44Þ
N i¼1 ð2pÞd=2 kd detðSÞ1=2 2k2

where Xi is the ith multivariate data point for a sample of size N, S is the sample
covariance of the variable set X.
The bandwidth λ is key to an accurate estimate of the probability density. A large
value of λ results in an oversmoothed probability density, with subdued models and
overenhanced tails. A small value, on the other hand, can lead to density estimates
overly influenced by individual data points, with noticeable bumps in the tails of the
probability density. Several operational rules for choosing optimal values of λ are
available in the literature; see Silverman (1986), Adamowski and Feluch (1991),
Scott (1992), Sain et al. (1994), and Rajagopalan et al. (1997) for some details. One
of the widely used guidelines is the least squares cross validation (LSCV).
The kernel density estimation-based approach started finding its applications in
hydrology in the 1980s; see Adamowski (1985), Schuster and Yakowitz (1985),
Bardsley (1989), Adamowski and Feluch (1990), Guo (1991), Lall et al. (1993,
1996), Moon and Lall (1994), Lall (1995), and Rajagopalan and Lall (1995) for
some early studies. However, the real impetus came in the latter part of the 1990s.
Sharma et al. (1997) presented a multivariate kernel density estimate as a non-
parametric alternative to the lag-p autoregressive model for monthly streamflow
simulation. Their model, called the NPp model, specifically considered the lag-1
situation (i.e. p = 1), (i.e. NP1 model), which means that the problem is bivariate
kernel density estimate. They used kernel density estimators with Gaussian kernels,
and selected the bandwidth λ using least squares cross validation (LSCV). Their
procedure for nonparametric streamflow simulation is as follows: (1) Form bivariate
sample set Xi = (Xi, Xi–1); (2) Estimate bandwidth λ using least squares cross
validation and estimate S; (3) Initialize i = 0, and obtain xi=0 (The initialization can
be done in two ways, either by sampling from the marginal or by using warm-up);
(4) From the given value xi–1 = Xi–1, select one of the observations Xi according to
weight wi; (5) Simulate xi from ith kernel slice, and correct for negative simulations;
and (6) Repeat steps (1) to (5) until the desired length of data is simulated. Further
92 3 Stochastic Time Series Methods

details about the procedure, including the use of the LSCV method for bandwidth
selection and its sensitivity analysis for streamflow simulation, are available in
Sharma et al. (1998).
The nonparametric kernel density estimator by Sharma et al. (1997) has been
modified to incorporate other characteristics of hydrologic time series. For instance,
Sharma and O’Neill (2002) developed a model to incorporate the long-term
interannual variability for monthly streamflow simulation (NPL model). They
particularly focused on order 1, and called the model as NPL1 model. Salas and Lee
(2010) suggested other modifications. In the development of a k-nearest neighbor
resampling algorithm with gamma kernel perturbation model (KGK model), they
used an aggregate variable (KGKA model) and a pilot variable (KGKP model) to
the kernel density estimator to lead the generation of the seasonal flows. They
compared the proposed models with another nonparametric model that considers
the reproduction of the interannual variability.
Kernel density estimator-based nonparametric models have found a number of
applications in hydrology over the last two decades, including for rainfall and
streamflow simulation and disaggregation (or downscaling). Extensive details of
such applications are available in Rajagopalan et al. (1997), Tarboton et al. (1998),
Rajagopalan and Lall (1999), Harrold et al. (2003a, b), Kim and Valdés (2005),
Srikanthan et al. (2005), Ghosh and Mujumdar (2007), Block et al. (2009),
Mehrotra and Sharma (2007a, b, 2010), Mehrotra et al. (2012), Li et al. (2013),
Mirhosseini et al. (2015), and Viola et al. (2016), among others.
Although the kernel-based methods have been found to be useful for a number
of hydrologic applications, they also possess certain limitations. For instance, they
often have problems at the boundaries of the variables (e.g. zero in rainfall and
streamflow), thus resulting in bias (e.g. Lall and Sharma 1996), which gets exag-
gerated at higher dimensions. The methods can also simulate negative values,
although this is to a lesser degree when compared to parametric models. The
methods are also very difficult to use in higher dimensions, since estimation of
optimal bandwidths at such dimensions is not trivial. These limitations led to other
nonparametric methods. Among these, the methods based on the concept of nearest
neighbors have been of particular significance and, therefore, are discussed next.

3.6.3 k-Nearest Neighbor Resampling (KNNR)

In the k-nearest neighbor resampling (e.g. Lall and Sharma 1996), the conditional
PDF is approximated using k-nearest neighbors of the current value Xi in the time
series and one of the neighbors is selected as the value for the next time
step. A brief description of the k-nearest neighbor resampling method for a single
variable time series is as follows:
3.6 Nonparametric Methods 93

1. Formulate a conditioning ‘feature vector’ based on the current value Xi of the

time series and the past values (i.e. Xi–1, Xi–2, …), with the number of past
values identified based on serial correlation;
2. Identify the k nearest neighbors of the feature vector based on Euclidean dis-
tance metric;
3. Assign weights to these k neighbors based on their distances, with the nearest
neighbor getting the largest weight and the farthest neighbor getting the smallest
weight;
4. Normalize these weights to create a probability mass function. This is termed as
the ‘weight metric;’
5. Resample one of the neighbors using this ‘weight metric;’
6. The successor of the resampled neighbor becomes the simulated value for the
next time step, Xi+1.
7. Repeat Steps (1)–(6) to generate several simulations.
This approach is simple and robust, as it can easily simulate from conditional
PDF of any dimension. The simulations are generally insensitive to the choice of
the weight function, as long as the selected weight function weighs the nearest
neighbor the most relative to the farthest.
The number of neighbors k can be thought of as a smoothing parameter. For
smaller k, the PDF approximation is based on only a few points and, therefore, has
the ability to capture the local features. A larger k, on the other hand, can smooth
out local features, but can capture global features. Therefore, an appropriate
selection of k is key to the performance of the k-nearest neighbor resampling
method. There are several methods for selecting an appropriate number for k,
including heuristic and objective methods (e.g. Generalized Cross Validation
pffiffiffiffi
(GCV)). In general, the choice k ¼ N is quite robust.
Although k-NN technique has been found to perform reasonably well for a
number of applications in hydrology, it has an important drawback in that the
values not seen in the historical record cannot be simulated. This means, there may
not be enough variety in simulations, while the use of kernel density estimators can
alleviate this problem (e.g. Sharma et al. 1997; Tarboton et al. 1998). Another
shortcoming is that the variance may be underestimated, especially when the his-
torical time series are correlated. This problem may be overcome by using gamma
kernel perturbation, as has been done by Salas and Lee (2010), mentioned earlier.
The nonparametric k-NN method and their modifications (see also Sect. 3.6.4 for
another example), including for multiple sites and multiple variables have found
numerous applications in hydrology. For instance, Rajagopalan and Lall (1999)
implemented a lag-1 multivariate resampling model to simulate daily rainfall and
other weather variables at a single site. Yates et al. (2003) extended this model to
simulate daily rainfall and weather variables at multiple sites simultaneously.
A semi-parametric approach to rainfall and weather generation was proposed. In
this, a Markov chain model was fitted for each month separately to simulate the
94 3 Stochastic Time Series Methods

rainfall occurrence and then k-NN lag-1 resampling was used to generate the vector
of weather variables conditionally on the transitional state of the rainfall occurrence.
For instance, if the simulated current period’s occurrence was wet and the following
period dry, then the neighbors are obtained from the historical time series that have
the same transition (Apipattanavis et al. 2007). Souza Filho and Lall (2003) applied
the k-NN method for multisite seasonal ensemble streamflow forecasts by formu-
lating the feature vector as a set of predictors for streamflow. Still other studies that
have applied the k-NN method in hydrology include those by Karlsson and
Yakowitz (1987), Buishand and Brandsma (2001), Brath et al. (2002), Mehrotra
and Sharma (2006), Prairie et al. (2007), Towler et al. (2009), Eum et al. (2010),
Gangopadhyay et al. (2005, 2009), Goyal et al. (2012), Lee and Jeong (2014), Lu
and Xin (2014), and Sharifazari and Araghinejad (2015), among others.

3.6.4 k-Nearest Neighbors with Local Polynomial

Regression

As mentioned earlier, the k-NN bootstrap method has an important limitation in that
it cannot simulate the values that are not seen in the historical record. This limi-
tation can be addressed by making certain modifications to the method. One such
modification for streamflow simulation was presented by Prairie et al. (2006) based
on local polynomial regression (Loader 1999). In this method, with the value of the
current period available, local regression is used to obtain the mean value of a future
period. Then, k neighbors are computed from the data for the current period and
residuals from the regression at these k neighbors are resampled using the k-NN
approach (see Sect. 3.6.3) and added to the mean value. Therefore, in this method,
instead of resampling the historical values, residuals are resampled from the
neighborhood. A brief description of this method is as follows (Prairie et al. 2006):
1. Fit a local polynomial for each timestep i dependent on the previous timestep
i−1

Xi ¼ gðXi1 Þ þ ei ð3:45Þ

where g(Xi–1) is the local polynomial;

2. Save the residuals (εi) from the fit;

3. Once the simulated value of the current period (i.e. Xi1 ) is obtained, estimate
b 
the mean flow of the next period X i from Eq. (3.45), not including the residual;

4. Obtain k-NN of Xi1 ;
5. Assign weights to these k neighbors based on their distances, with the nearest
neighbor getting the largest weight and the farthest neighbor getting the
smallest;
3.6 Nonparametric Methods 95

6. Normalize these weights to create a probability mass function. This is termed as

the ‘weight metric;’  
7. Resample one of the neighbors using this ‘weight metric.’ Its residual ei is
added to the mean estimate Xb i . Thus, the simulated value for the next timestep
becomes Xi ¼ Xb i þ ei ;
8. Repeat Steps (1)–(7) for other time periods to obtain an ensemble of
simulations.
The local polynomial based k-NN method has the following advantages over the
traditional k-NN method: (1) it provides simulations even for values not seen in the
historical record; (2) residual resampling captures the local variability more effec-
tively; and (3) the local regression fit has the ability to capture any arbitrary (linear
or nonlinear) relationship in the time series and also extrapolate beyond the range of
observations (Prairie et al. 2006).
The local polynomial k-NN method and its further extensions has been suc-
cessfully applied to study different problems in hydrology, including for rainfall and
streamflow simulation, forecasting, and downscaling. Extensive details of such
applications can be seen in Regonda et al. (2005, 2006a, b), Singhrattna et al.
(2005), Grantz et al. (2005), Block and Rajagopalan (2007), Bracken et al. (2010),
Li and Singh (2014), and Verdin et al. (2016), among others.

3.6.5 Others

There exist many other stochastic nonparametric methods that have also found
important applications in hydrology. In some cases, nonparametric models have
been combined with parametric models to serve as hybrid models. For instance,
Srinivas and Srinivasan (2001) proposed a hybrid moving block bootstrap (HMBB)
model for multi-season streamflow simulation, by combining a parametric periodic
model and a nonparametric block bootstrap. Srinivas and Srinivasan (2005, 2006)
presented further modifications and improvements to this hybrid model. Srivatsav
et al. (2011) used the hybrid model of Srinivas and Srinivasan (2006) as a simu-
lation engine to develop an efficient simulation-optimization-based hybrid
stochastic modeling framework. Srivatsav and Simonovic (2014) used the maxi-
mum entropy bootstrap model to provide analytical procedures for multi-site,
multi-season weather generation and streamflow generation; see also Cook and
Buckley (2009), and Cook et al. (2013) for maximum entropy bootstrap model
applications in hydrology. Still other recent studies of interest on nonparametric or
hybrid models in hydrology are Lambert et al. (2003), Mehrotra and Sharma
(2007a), Wong et al. (2007), Basinger et al. (2010), Lee et al. (2010), Kalra and
Ahmad (2011), Keylock (2012), Stagge and Moglen (2013), Haerter et al. (2015),
Mehrotra et al. (2015), and Langousis et al. (2016), among others.
96 3 Stochastic Time Series Methods

3.7 Summary

Over the past half a century or so, stochastic time series methods have been an
important part of hydrologic studies. A large number of parametric and nonpara-
metric (as well as hybrid or semi-parametric) methods have been developed and
applied in almost all areas of hydrology, including for simulation, forecasting, and
disaggregation (or downscaling) of rainfall and streamflow. This chapter has pre-
sented an overview of several parametric and nonparametric methods and their
hydrologic applications. It is clear that stochastic time series methods have domi-
nated hydrologic studies in recent decades, and there is no question that this trend
will continue to grow in the future. In this regard, an area that is gaining particular
attention is the application of stochastic methods for downscaling outputs from
global climate models (GCMs), especially with the clear recognition of the impacts
of climate change on our water resources at the global scale as well as at the
regional and local scales. While stochastic time series methods have been domi-
nating hydrologic research in recent decades, a large number of other approaches,
especially those that address the nonlinear and related properties of time series,
have also been developed and extensively applied in hydrology. Some of these
methods that have found widespread applications in hydrology are discussed in the
next chapter.

References

Adamowski K (1985) Nonparametric kernel estimation of flood frequencies. Water Resour Res 21
(11):1585–1590
Adamowski K, Feluch W (1990) Nonparametric flood-frequency analysis with historical
information. J Hydraul Eng 116(8):1035–1047
Adamowski K, Feluch W (1991) Application of nonparametric regression to ground water level
prediction. Can J Civ Eng 18:600–606
Ahn H, Salas JD (1997) Groundwater head sampling based on stochastic analysis. Water Resour
Res 33(12):2769–2780
Apipattanavis S, Podestá G, Rajagopalan B, Katz RW (2007) A semiparametric multivariate and
multisite weather generator. Water Resour Res 43:W11401. doi:10.1029/2006WR005714
Aykuz DE, Bayazit M, Önöz B (2012) Markov chain models for hydrological drought
characteristics. J Hydrometeorol 13(1):298–309
Bardossy A, Plate EJ (1991) Space-time model for daily rainfall using atmospheric circulation
patterns. Water Resour Res 28(5):1247–1259
Bardsley WE (1989) Using historical data in nonparametric flood estimation. J Hydrol 108:249–
255
Bartlett MS (1955) An introduction to stochastic processes. Cambridge University Press,
Cambridge
Bartlett MS (1963) The spectral analysis of point processes. J R Stat Soc Ser B 25(2):264–296
Bartolini P, Salas JD (1993) Modeling of streamflow processes at different time scales. Water
Resour Res 29(8):2573–2587
Bartolini P, Salas JD, Obeysekera JTB (1988) Multivariate periodic ARMA(1,1) processes. Water
Resour Res 24(8):1237–1246
References 97

Basinger M, Montalto F, Lall U (2010) A rainwater harvesting system reliability model based on
nonparametric stochastic rainfall generator. J Hydrol 392:105–118
Bates BC, Campbell EP (2001) A Markov chain Monte Carlo scheme for parameter estimation and
inference in conceptual rainfall-runoff modeling. Water Resour Res 37(4):937–947
Bayazit M, Önöz B (2005) Probabilities and return periods of multisite droughts. Hydrol Sci J
50:605–615
Berndstein S (1938) Equations différentielles stochastiques. Actualités Sci Ind No 138:5–31
Block P, Rajagopalan B (2007) Interannual variability and ensemble forecast of Upper Blue Nile
Basin Kiremt season precipitation. J Hydrometeorol 8:327–343
Block PJ, Souza Filho FA, Sun L, Kwon H-H (2009) A streamflow forecasting framework using
multiple climate and hydrological models. J Ame Wat Resour Assoc 45(4):828–843
Bo Z, Islam S, Eltahir EAB (1994) Aggregation-disaggregation properties of a stochastic rainfall
model. Water Resour Res 30(12):3423–3435
Bochner S (1949) Diffusion equations and stochastic processes. Proc Natl Acad Sci USA 35
(7):368–370
Bochner S (1955) Harmonic analysis and the theory of probability. University of California Press,
Berkeley
Bordoy R, Burlando P (2014) Stochastic downscaling of climate model precipitation outputs in
orographically complex regions: 2. Downscaling methodology. Water Resour Res 50(1):562–
579
Box G, Jenkins G (1970) Time series analysis, forecasting, and control. Holden-Day, San
Francisco
Bracken C, Rajagopalan B, Prairie J (2010) A multisite seasonal ensemble streamflow forecasting
technique. Water Resour Res 46(3):W03532. doi:10.1029/2009WR007965
Bras RL, Rodriguez-Iturbe I (1985) Random functions and hydrology. Addison-Wesley, MA
Brath A, Montanari A, Toth E (2002) Neural networks and non-parametric methods for improving
real-time forecasting through conceptual hydrological models. Hydrol Earth Syst Sci 6(4):627–
640
Brillinger DR (1975) Time series: data analysis and theory. Holt, Rinehart & Winston, New York
Brockwell PJ, Davis RA (1991) Time series: theory and methods. Springer-Verlag, New York
Buishand TA (1977) Stochastic modeling of daily rainfall sequences. Veenman and Zonen,
Wageningen, The Netherlands
Buishand TA, Brandsma T (2001) Multisite simulation of daily precipitation and temperature in
the Rhine basin by nearest-neighbor resampling. Water Resour Res 37(11):2761–2776
Burton A, Fowler HJ, Blenkinsop S, Kilsby CG (2010) Downscaling transient climate change
using a Neyman-Scott Rectangular Pulses stochastic rainfall model. J Hydrol 381:18–32
Calenda G, Napolitano F (1999) Parameter estimation of Neyman-Scott processes for temporal
point rainfall simulation. J Hydrol 225:45–66
Cameron D, Beven KJ, Tawn J (2000) Modelling extreme rainfalls using a modified random pulse
Bartlett-Lewis stochastic rainfall model (with uncertainty). Adv Water Resour 24:203–211
Cancelliere A, Salas JD (2004) Drought length probabilities for periodic-stochastic hydrologic
data. Water Resour Res 40:W02503. doi:10.1029/2002WR001750
Cancelliere A, Salas JD (2010) Drought probabilities and return period for annual streamflow
series. J Hydrol 391:77–89
Carlson RF, MacCormick AJA, Watts DG (1970) Application of linear models to four annual
streamflow series. Water Resour Res 6:1070–1078
Carsteanu A, Foufoula-Georgiou E (1996) Assessing dependence of weights in a multiplicative
cascade model of temporal rainfall. J Geophys Res 101:26363–26370
Chaleeraktrakoon (1999) Stochastic procedure for generating seasonal flows. J Hydrol Eng 4
(4):337–343
Chaleeraktrakoon C (2009) Parsimonious SVD/MAR(1) procedure for generating multisite
multiseason flows. J Hydrol Eng 14(5):516–527
Chang TJ, Kavvas ML, Delleur JW (1984) Daily precipitation modeling by discrete autoregressive
moving average processes. Water Resour Res 20(5):565–580
98 3 Stochastic Time Series Methods

Chatfield C (1989) The analysis of time series: An introduction. CRC Press, USA
Chin EH (1977) Modeling daily precipitation occurrence process with Markov chains. Water
Resour Res 13(6):849–956
Chung C, Salas JD (2000) Drought occurrence probabilities and risk of dependent hydrological
process. J Hydrol Eng 5:259–268
Cigizoglu HK, Bayazit M (1998) Application of gamma autoregressive model to analysis of dry
periods. J Hydrol Eng 3(3):218–221
Claps P, Rossi F, Vitale C (1993) Conceptual-stochastic modeling of seasonal runoff using
autoregressive moving average models and different time scales of aggregation. Water Resour
Res 29(8):2545–2559
Clarke RT (1994) Statistical modelling in hydrology. John Wiley & Sons, Chichester, UK
Connolly RD, Schirmer J, Dunn PK (1998) A daily rainfall disaggregation model. Agric For
Meteorol 92:105–117
Cook BI, Buckley BM (2009) Objective determination of monsoon season onset, withdrawal, and
length. J Geophys Res 114(D23):D23109. doi:10.1029/2009JD012795
Cook ER, Palmer JG, Ahmed M, Woodhouse CA, Fenwick P, Zafar MU, Wahab M, Khan N
(2013) Five centuries of Upper Indus River flow from tree rings. J Hydrol 486:365–375
Cooper DM, Wood EF (1982) Identification of multivariate time series and multivariate
input-output models. Water Resour Res 18(4):937–946
Cowpertwait PSP (1991) Further developments of the Neyman-Scott clustered point process for
modeling rainfall. Water Resour Res 27:1431–1438
Cowpertwait PSP (1995) A generalized spatial-temporal model of rainfall based on a clustered
point process. Proc R Soc London Ser A 450:163–175
Cowpertwait PSP (1998) A Poisson-cluster model of rainfall: high-order moments and extreme
values. Proc R Soc London Ser A 454:885–898
Cowpertwait PSP (2004) Mixed rectangular pulses models of rainfall. Hydrol Earth Syst Sci 8
(5):993–1000
Cowpertwait PSP (2010) A spatial-temporal point process model with a continuous distribution of
storm types. Water Resour Res 46(12):W12507. doi:10.1029/2010WR009728
Cowpertwait PSP, O’Connell PE (1997) A regionalised Neyman-Scott model of rainfall with
convective and stratiform cells. Hydrol Earth Syst Sci 1:71–80
Cowpertwait PSP, O’Connell PE, Metcalfe AV, Mawdsley J (1996a) Stochastic point process
modelling of rainfall, I. Single-site fitting and validation. J Hydrol 175(1–4):17–46
Cowpertwait PSP, O’Connell PE, Metcalfe AV, Mawdsley J (1996b) Stochastic point process
modelling of rainfall, II. Regionalisation and disaggregation. J Hydrol 175(1–4):47–65
Cowpertwait PSP, Kilsby CG, O’Connell PE (2002) A space-time Neyman-Scott model of
rainfall: empirical analysis of extremes. Water Resour Res 38(8):1–14
Cowpertwait PSP, Isham V, Onof C (2007) Point process models of rainfall: developments for
fine-scale structure. Proc R Society London Ser A 463(2086):2569–2587
Cowpertwait PSP, Xie G, Isham V, Onof C, Walsh DCI (2011) A fine-scale point process model
of rainfall with dependent pulse depths within cells. Hydrol Sci J 56(7):1110–1117
Cox DR (1958) Some problems connected with statistical inference. Ann Math Statist 29(2):357–
372
D’Onofrio D, Palazzi E, von Hardenberg J, Provenzale A, Calmanti S (2014) Statistical rainfall
downscaling of climate models. J Hydrometeorol 15:830–843
Deidda R (2000) Rainfall downscaling in a space-time multifractal framework. Water Resour Res
36(7):1779–1784
Delleur JW, Kavvas ML (1978) Stochastic models for monthly rainfall forecasting and synthetic
generation. J Appl Meteorol 17:1528–1536
Delleur JW, Tao PC, Kavvas ML (1976) An evaluation of the practicality and complexity of some
rainfall and runoff time series models. Water Resour Res 12(5):953–970
Dodov B, Foufoula-Georgiou E (2005) Fluvial processes and streamflow variability: Interplay in
the scale-frequency continuum and implications for scaling. Water Resour Res 41:W05005.
doi:10.1029/2004WR003408
References 99

Doob JL (1934) Stochastic processes and statistics. Proc Natl Acad Sci USA 20:376–379
Doob JL (1938) Stochastic processes with an integer-valued parameter. Trans Amer Math Soc
44:87–150
Doob JL (1940) Regularity properties of certain families of chance variables. Trans Amer Math
Soc 47:455–486
Doob JL (1945) Markov chains–denumerable case. Trans Amer Math Soc 58:455–473
Doob JL (1953) Stochastic processes. John Wiley & Sons Inc, New York
Dynkin EB (1959) Foundations of the theory of Markov processes. Fizmatgiz, Moscow
Efron B (1979) Bootstrap methods: Another look at the Jackknife. Ann Stat 7:1–26
Efron B, Tibishirani R (1993) An introduction to the bootstrap. Chapman and Hall, New York
Entekhabi D, Rodriguez-Iturbe I, Eagleson PS (1989) Probabilistic representation of the temporal
rainfall process by the modified Neyman-Scott rectangular pulses model: parameter estimation
and validation. Water Resour Res 25(2):295–302
Eubank RL (1988) Spline smoothing and nonparametric regression. Marcel Dekker Inc, New York
Eum H, Simonovic S, Kim Y (2010) Climate change impact assessment using k-nearest neighbor
weather generator: case study of the Nakdong River Basin in Korea. J Hydrol Eng 15(10):772–
785
Evin G, Favre A-C (2008) A new rainfall model based on the Neyman-Scott process using cubic
copulas. Water Resour Res 44:W03433. doi:10.1029/2007WR006054
Favre A-C, Musy A, Morgenthaler S (2002) Two-site modeling of rainfall based on the
Neyman-Scott process. Water Resour Res 38(12):1307. doi:10.1029/2002WR001343
Feller W (1950) An introduction to probability theory and its applications. John Wiley, New York
Fernandez B, Salas JD (1986) Periodic gamma-autoregressive processes for operational
hydrology. Water Resour Res 22(10):1385–1396
Fernandez B, Salas JD (1990) Gamma-autoregressive models for stream-flow simulation.
J Hydraul Eng 116(11):1403–1414
Ferraris L, Gabellani S, Parodi U, Rebora N, von Hardenberg J, Provenzale A (2003) Revisiting
multifractality in rainfall fields. J Hydrometeorol 4:544–551
Fiering M (1967) Streamflow synthesis. Hard University Press, Cambridge, MA
Fiering MB, Jackson BB (1971) Synthetic streamflows. Am Geophys Union Water Resour,
Monograph 1:1–98
Foufoula-Georgiou E, Guttorp P (1986) Compatibility of continuous rainfall occurrence models
with discrete rainfall observations. Water Resour Res 22(8):1316–1322
Foufoula-Georgiou E, Lettenmaier DP (1987) A Markov renewal model of rainfall occurrences.
Water Resour Res 23(5):875–884
Fowler HJ, Blenkinsop S, Tebaldi C (2007) Linking climate change modeling to impacts studies:
recent advances in downscaling techniques for hydrological modeling. Int J Climatol 27
(12):1547–1578
Fraedrich K, Larnder C (1993) Scaling regimes of composite rainfall time series. Tellus 45A:289–
298
Gabriel KR, Neumann J (1962) A Markov chain model for daily rainfall occurrence at Tel Aviv.
Q J R Meteorol Soc 88:90–95
Gangopadhyay S, Clark M, Rajagopalan B (2005) Statistical downscaling using K-nearest
neighbors. Water Resour Res 41:W02024. doi:10.1029/2004WR003444
Gangopadhyay S, Harding BL, Rajagopalan B, Lukas JJ, Fulp TJ (2009) A nonparametric
approach for paleohydrologic reconstruction of annual streamflow ensembles. Water Resour
Res 45:W06417. doi:10.1029/2008WR007201
Gani J (1969) Recent advances in storage and flooding theory. Adv Appl Prob 1:90–110
Gardiner CW (2009) Stochastic methods: a handbook for the natural and social sciences.
Springer-Verlag, Berlin and Heidelberg
Gaver DP, Lewis PAW (1980) First-order autoregressive gamma sequences and point processes.
Adv Appl Prob 12:727–745
Gelhar LW (1993) Stochastic subsurface hydrology. Prentice-Hall, Englewood Cliffs, New Jersey
100 3 Stochastic Time Series Methods

Ghosh S, Mujumdar PP (2007) Nonparametric methods for modeling GCM and scenario
uncertainty in drougth assessment. Water Resour Res 43:W07405. doi:10.1029/
2006WR005351
Glasbey CA, Cooper C, McGechan MB (1995) Disaggregation of daily rainfall by conditional
simulation from a point-process model. J Hydrol 165:1–9
Govindaraju RS (ed) (2002) Stochastic methods in subsurface contaminant hydrology. ASCE,
New York
Goyal MK, Ojha CSP, Burn DH (2012) Nonparametric statistical downscaling of temperature,
precipitation, and evaporation in a semiarid region in India. J Hydrol Eng 17(5):615–627
Grantz K, Rajagopalan B, Clark M, Zagona E (2005) A technique for incorporating large-scale
climate information in basin-scale ensemble streamflow forecasts. Water Resour Res 41:
W10410. doi:10.1029/2004WR003467
Grillakis MG, Koutroulis AG, Tsanis IK (2013) Multisegment statistical bias correction of daily
GCM precipitation output. J Geophys Res 118(8):3150–3162
Grygier JC, Stedinger JR (1988) Condensed disaggregation procedures and conservation
corrections for stochastic hydrology. Water Resour Res 24(10):1574–1584
Guo SL (1991) Nonparametric variable kernel estimation with historical floods and paleoflood
information. Water Resour Res 27(1):91–98
Gupta VK, Waymire E (1981) The mathematical structure of rainfall representations: 3. Some
applications of the point process theory to rainfall processes. Water Resour Res 17(5):1287–
1294
Gupta VK, Waymire E (1990) Multiscaling properties of spatial rainfall and river flow
distributions. J Geophys Res 95(D3):1999–2009
Gupta VK, Waymire E (1993) A statistical analysis of mesoscale rainfall as a random cascade.
J Appl Meteorol 32:251–267
Gupta VK, Rodríguez-Iturbe I, Wood EF (eds) (1986) Scale problems in hydrology: runoff
generation and basin response. Reidel Publishing Company, FD, p 244
Gupta VK, Mesa OJ, Dawdy DR (1994) Multiscaling theory of flood peaks: Regional quantile
analyis. Water Resour Res 30(12):3405–3421
Gupta VK, Castro S, Over TM (1996) On scaling exponents of spatial peak flows from rainfall and
river network geometry. J Hydrol 187(1–2):81–104
Gyasi-Agyei Y, Willgoose G (1997) A hybrid model for point rainfall modeling. Water Resour
Res 33:1699–1706
Haan CT (1994) Statistical methods in hydrology. Iowa University Press, Iowa
Haan CT, Allen DM, Street JD (1976) A Markov chain model of daily rainfall. Water Resour Res
12(3):443–449
Haerter JO, Eggert B, Moseley C, Piani C, Berg P (2015) Statistical precipitation bias correction of
gridded model data using point measurements. Geophys Res Lett 42:1919–1929. doi:10.1002/
2015GL063188
Hannan EJ (1955) A test for singularities in Sydney rainfall. Austr Jour Phys 8(2):289–297
Härdle W, Bowman AW (1988) Bootstrapping in nonparametric regression: local adaptive
smoothing and confidence bands. J Am Stat Assoc 83:102–110
Harms AA, Campbell TH (1967) An extension to the Thomas-Fiering model for the sequential
generation of streamflow. Water Resour Res 3(3):653–661
Harrold TI, Sharma A, Sheather SJ (2003a) A nonparametric model for stochastic generation of
daily rainfall occurrence. Water Resour Res 39(10):1300. doi:10.1029/2003WR002182
Harrold TI, Sharma A, Sheather SJ (2003b) A nonparametric model for stochastic generation of
daily rainfall amounts. Water Resour Res 39(12):1343. doi:10.1029/2003WR002570
Hershenhorn J, Woolhiser DA (1987) Disaggregation of daily rainfall. J Hydrol 95:299–322
Higgins JJ (2003) Introduction to modern nonparametric statistics. Duxbury Press, North Scituate,
Massachusetts
Hipel KW, McLeod AI (1978a) Preservation of the rescaled adjusted range. 2, Simulation studies
using Box-Jenkins models. Water Resour Res 14(3):509–516
References 101

Hipel KW, McLeod AI (1978b) Preservation of the rescaled adjusted range. 3, Fractional Gaussian
noise algorithms. Water Resour Res 14(3):517–518
Hipel KW, McLeod AI (1994) Time series modeling of water resources and environmental
systems. Elsevier Science, Amsterdam
Hipel KW, McLeod AI, Lennox WC (1977) Advances in Box-Jenkins modeling, 1 Model
construction. Water Resour Res 13(3):567–575
Hirsch RM (1979) Synthetic hydrology and water supply reliability. Water Resour Res 15
(6):1603–1615
Hirsch RM, Archfield SA, De Cicco LA (2015) A bootstrap method for estimating uncertainty of
water quality trends. Environ Modell Softw 73:148–166
Hollander M, Wolfe DA, Chicken E (2013) Nonparametric statistical methods. John Wiley &
Sons, Hoboken, New Jersey
Hoshi K, Burges SJ (1979) Disaggregatio of streamflow volumes. J Hydraul Div ASCE 195
(HY1):27–41
Hughes J, Guttorp P, Charles S (1999) A non-homogeneous hidden Markov model for
precipitation occurrence. Appl Stat 48(1):15–30
Hurst HE (1951) Long-term storage capacity of reservoirs. Trans Am Soc Civil Eng 116:770–808
Hurst HE (1956) Methods of using long-term storage in reservoirs. Proc Inst Civil Eng 1:519–543
Islam S, Entekhabi D, Bras RL, Rodriguez-Iturbe I (1990) Parameter estimation and sensitivity
analysis for the modified Bartlett-Lewis rectangular pulses model of rainfall. J Geophys Res
95:2093–2100
Itô K (1944) Stochastic integral. Proc Imp Acad Tokyo 20:519–524
Itô (1946) On a stochastic integral equation. Proc Japan Acad Tokyo 22:32–35
Jackson BB (1975) Markov mixture models for drought lengths. Water Resour Res 11(1):75–95
Jazwinski A (1970) Stochastic processes and filtering theory. Academic Press, New York
Jenkins GM, Watts DG (1968) Spectral analysis and its applications. Holden-Day, San Francisco
Kaczmarska J, Isham V, Onof C (2014) Point process models for finite-resolution rainfall. Hydrol
Sci J 59(11):1972–1991
Kalra A, Ahmad S (2011) Evaluating changes and estimating seasonal precipitation for Colorado
River Basin using stochastic nonparametric disaggregation technique. Water Resour Res 47:
W05555. doi:10.1029/2010WR009118
Karlin S, Taylor HM (1975) A first course in stochastic processes. Academic Press, San Diego,
California
Karlsson M, Yakowitz S (1987) Nearest-neighbor methods for nonparametric rainfall-runoff
forecasting. Water Resour Res 23(7):1300–1308
Katz RW (1977) Precipitation as a chain-dependent process. J Appl Meteor 16:671–676
Katz RW, Parlange MB (1995) Generalizations of chain-dependent processes: application to
hourly precipitation. Water Resour Res 31:1331–1341
Kavvas ML, Delleur JW (1981) A stochastic cluster model of daily rainfall sequences. Water
Resour Res 17(4):1151–1160
Kendall DR, Dracup JA (1991) A comparison of index-sequential and AR(1) generated
hydrological sequences. J Hydrol 122:335–352
Kendall MG, Stuart A (1968) Advanced theory of statistics. Wiley, New York
Keylock KJ (2012) A resampling method for generating synthetic hydrological time series with
preservation of cross-correlative structure and higher-order properties. Water Resour Res 48:
W12521. doi:10.1029/2012WR011923
Khaliq MN, Cunnane C (1996) Modelling point rainfall occurrences with the modified
Bartlett-Lewis rectangular pulses model. J Hydrol 180(1–4):109–138
Kim T, Valdés B (2005) Synthetic generation of hydrologic time series based on nonparametric
random generation. J Hydrol Eng 10(5):395–404
Klemeš V (1978) Physically based stochastic hydrologic analysis. Adv Hydrosci 11:285–352
Kottegoda NT (1980) Stochastic water resources technology. The Macmillan Press, London
Koutsoyiannis D (1992) A nonlinear disaggregation method with a reduced parameter set for
simulation of hydrologic series. Water Resour Res 28(12):3175–3191
102 3 Stochastic Time Series Methods

Koutsoyiannis D (2000) A generalized mathematical framework for stochastic simulation and

forecast of hydrologic time series. Water Resour Res 36:1519–1534
Koutsoyiannis D (2001) Coupling stochastic models of different timescales. Water Resour Res 37
(2):379–391
Koutsoyiannis D, Onof C (2001) Rainfall disaggregation using adjusting procedures on a Poisson
cluster model. J Hydrol 246(1–4):109–122
Koutsoyiannis D, Xanthopoulos Th (1990) A dynamic model for short-scale rainfall disaggre-
gation. Hydrol Sci J 35(3):303–322
Kundzewicz ZW, Robson AJ (2004) Change detection in hydrological records—a review of the
methodology. Hydrol Sci J 49(1):7–19
Künsch HR (1989) The jackknife and the bootstrap for general stationary observations. Ann Stat
17:1217–1241
Labadie JW, Fontane D, Tabios GQ, Chou NF (1987) Stochastic analysis of dependable
hydropower capacity. J Water Resour Plann Manage 113(3):422–477
Lall U (1995) Nonparametric function estimation: recent hydrologic applications. US Natl Rep Int
Union Geod Geophys 1991–1994. Rev Geophys 33:1093–1102
Lall U, Sharma A (1996) A nearest neighbor bootstrap for resampling hydrologic time series.
Water Resour Res 32(3):679–693
Lall U, Moon Y-I, Bosworth K (1993) Kernel flood frequency estimators: bandwidth selection and
kernel choice. Water Resour Res 29(4):1003–1016
Lall U, Rajagopalan B, Tarboton DG (1996) A nonparametric wet/dry spell model for resampling
daily precipitation. Water Resour Res 32:2803–2823
Lambert MF, Whiting JP, Metcalfe AV (2003) A non-parametric hidden Markov model for
climate change identification. Hydrol Earth Syst Sci 7(5):652–667
Lane WL (1979) Applied stochastic techniques. User Manual. Division of Planning Technical
Services, US Bur Reclam, Denver, Colo
Langousis A, Mamalakis A, Puliga M, Deidda R (2016) Threshold detection for the generalized
Pareto distribution: Review of representative methods and application to the NOAA NCDC
daily rainfall database. Water Resour Res 52:2659–2681
Lawrance AJ (1982) The innovation distribution of a gamma distributed autoregressive process.
Scand J Statist 9:234–236
Lawrance AJ, Kottegoda NT (1977) Stochastic modeling of river flow time series. J R Stat Soc
Ser A 140:1–47
Lawrance AJ, Lewis PAW (1981) A new autoregressive time series model in exponential variables
[NEAR(1)]. Adv Appl Prob 13(4):826–845
Le Cam LA (1958) Un théorème sur la division d’un intervalle par des points pris au hasard. Publ
Inst Statist Univ Paris 7:7–16
Le Cam LA (1961) A stochastic description of precipitation. In: Newman J (ed) Proc 4th Berkeley
symp mathematics, statistics, and probability. University of California Press, Berkeley,
pp 165–186
Ledolter J (1978) The analysis of multivariate time series applied to problems in hydrology.
J Hydrol 36:327–352
Lee T, Jeong C (2014) Nonparametric statistical temporal downscaling of daily precipitation to
hourly precipitation and implications for climate change scenarios…
Lee T, Salas JD, Prairie J (2010) An enhanced nonparametric streamflow disaggregation model
with genetic algorithm. Water Resour Res 46:W08545. doi:10.1029/2009WR007761
Lennartsson J, Baxevani A, Chen D (2008) Modelling precipitation in Sweden using multiple step
Markov chains and a composite model. J Hydrol 363(1):42–59
Leonard M, Lambert MF, Metcalfe AV, Cowpertwait PSP (2008) A space-time Neyman-Scott
rainfall model with defined storm extent. Water Resour Res 44:W09402. doi:10.1029/
2007WR006110
Lettenmaier DP, Burges SJ (1977) Operational assessment of hydrologic models of long-term
persistence. Water Resour Res 13(1):113–124
Lévy P (1925) Calcul des probabilités. Gauthier-Villars, Paris
References 103

Lévy P (1937) Théorie de l’addition des variables aléatoires. Gauthier-Villars, Paris

Lévy (1948) Processus stochastiques et mouvement Brownien. Gauthier-Villars, Paris
Li C, Singh VP (2014) A multimodel regression-sampling algorithm for generating rich monthly
streamflow scenarios. Water Resour Res 50:5958–5979
Li C, Singh VP, Mishra AK (2013) A bivariate mixed distribution with a heavy-tailed component
and its application to single-site daily rainfall simulation. Water Resour Res 49:767–789
Lloyd EH (1963) A probability theory of reservoirs with serially correlated inputs. J Hydrol 1:99–
128
Loader CR (1999) Bandwidth selection: classical or plug-in? Ann Statist 27(2):415–438
Loève M (1955) Probability theory. D van Nostrand Company Inc, New York
Loucks DP, Stedinger JR, Haith DA (1981) Water resources systems planning and analysis.
Prentice-Hall, Englewood Cliffs, New Jersey
Lovejoy S, Schertzer D (1990) Multifractals, universality classes and satellite and radar
measurements of clouds and rain fields. J Geophys Res 95(D3):2021–2034
Lu Y, Xin X (2014) Multisite rainfall downscaling and disaggregation in a tropical urban area.
J Hydrol 509:55–65
Mandelbrot BB (1967) How long is the coast of Britain? Statistical self-similarity and fractional
dimension. Science 156:636–638
Mandelbrot BB (1972) Broken line process derived as an approximation to fractional noise. Water
Resour Res 8(5):1354–1356
Mandelbrot BB (1974) Intermittent turbulence in self-similar cascades: divergence of high
moments and dimension of the carrier. J Fluid Mech 62(2):331–358
Mandelbrot BB (1977) Fractals: form, chance and dimension. W.H. Freeman and Co, New York
Mandelbrot BB (1983) The fractal geometry of nature. Freeman, New York
Mandelbrot BB, Van Ness JW (1968) Fractional Brownian motions, fractional noises and
applications. SIAM Soc Ind Appl Math Rev 10(4):422–437
Mandelbrot BB, Wallis JR (1968) Noah, Joseph and operational hydrology. Water Resour Res 4
(5):909–918
Mandelbrot BB, Wallis JR (1969) Computer experiments with fractional Gaussian noises, 1.
Averages and variances. Water Resour Res 5(1):228–267
Mann HB (1945) Nonparametric tests against trend. Econometrica 13:245–259
Marani M, Zanetti S (2007) Downscaling rainfall temporal variability. Water Resour Res 43:
W09415. doi:10.1029/2006WR005505
Maraun D, Wetterhall F, Ireson AM, Chandler RE, Kendon EJ, Widmann M, Brienen S, Rust HW,
Sauter T, Themeßl M, Venema VKC, Chun KP, Goodess CM, Jones RG, Onof C, Vrac M,
Thiele-Eich I (2010) Precipitation downscaling under climate change: Recent developments to
bridge the gap between dynamical models and the end user. Rev Geophys 48:RG3003. doi:10.
1029/2009RG000314
Markonis Y, Koutsoyiannis D (2015) Scale-dependence of persistence in precipitation records.
Nature Climate Change 6:399–401
Marshall L, Nott D, Sharma A (2004) A comparative study of Markov chain Monte Carlo methods
for conceptual rainfall-runoff modeling. Water Resour Res 40:W02501. doi:10.1029/
2003WR002378
Matalas NC (1963a) Autocorrelation of rainfall and streamflow minimums. US Geol Surv Prof
Paper 434-B, Virginia
Matalas NC (1963b) Statistics of a runoff-precipitation relation. US Geol Surv Prof Paper 434-D,
Virginia
Matalas NC (1967) Mathematical assessment of synthetic hydrology. Water Resour Res 3(4):937–
945
Mathevet T, Lepiller M, Mangin A (2004) Application of time-series analyses to the hydrological
functioning of an Alpine karstic system: the case of Bange-L’Eau-Morte. Hydrol Earth Syst Sci
8(6):1051–1064
McCuen RH (2003) Hydrologic analysis and design. Prentice-Hall Inc., New Jersey
104 3 Stochastic Time Series Methods

McKean HP (1969) Stochastic integrals. American Mathematical Society, Academic Press, New
York
McKerchar AI, Delleur JW (1974) Application of seasonal parametric linear stochastic models to
monthly flow data. Water Resour Res 10(2):246–255
McLeod AI, Hipel KW, Lennox WC (1977) Advances in Box-Jenkins modeling. 2, Applications.
Water Resour Res 13(3):577–586
Mehrotra R, Sharma A (2006) Conditional resampling of hydrologic time series using multiple
predictor variables: A K-nearest neighbour approach. Adv Water Resour 29(7):987–999
Mehrotra R, Sharma A (2007a) A semi-parametric model for stochastic generation of multi-site
rainfall exhibiting low-frequency variability. J Hydrol 335:180–193
Mehrotra R, Sharma A (2007b) Preserving low-frequency variability in generated daily rainfall
sequences. J Hydrol 345:102–120
Mehrotra R, Sharma A (2010) Development and application of a multisite rainfall stochastic
downscaling framework for climate change impact assessment. Water Resour Res 46:W07526.
doi:10.1029/2009WR008423
Mehrotra R, Sharma A (2012) An improved standardization procedure to remove systematic low
frequency variability biases in GCM simulations. Water Resour Res 48:W12601. doi:10.1029/
WR012446
Mehrotra R, Westra S, Sharma A, Srikanthan R (2012) Continuous rainfall simulation: 2.
A regionalized daily rainfall generation approach. Water Resour Res 48:W01536. doi:10.1029/
2011WR010490
Mehrotra R, Li J, Westra S, Sharma A (2015) A programming tool to generate multi-site daily
rainfall using a two-stage semi parametric model. Environ Modell Softw 63:230–239
Mejia JM, Rouselle J (1976) Disaggregation models in hydrology revisited. Water Resour Res 12
(2):185–186
Mejia JM, Rodriguez-Iturbe I, Dawdy DR (1972) Streamflow simulation. 2, The broken line
process as a potential model for hydrologic simulation. Water Resour Res 8(4):931–941
Mejia JM, Dawdy DR, Nordin CF (1974) Streamflow simulation. 3, The broken line process and
operational hydrology. Water Resour Res 10(2):242–245
Menabde M, Harris D, Seed A, Austin G, Stow D (1997) Multiscaling properties of rainfall and
bounded random cascades. Water Resour Res 33(12):2823–2830
Meneveau C, Sreenivasan KR (1987) Simple multifractal cascade model for fully developed
turbulence. Phys Rev Lett 59:1424–1427
Mimikou (1984) A study for improving precipitation occurrences modelling with a Markov chain.
J Hydrol 70(1–4):25–33
Mirhosseini G, Srivastava P, Sharifi A (2015) Developing probability-based IDF curves using
kernel density estimator. J Hydrol Eng 20(9):04015002. doi:10.1061/(ASCE)HE.1943-5584.
0001160
Molnar P, Burlando P (2005) Preservation of rainfall properties in stochastic disaggregation by a
simple random cascade model. Atmos Res 77(1–4):137–151
Montanari A, Rosso R, Taqqu MS (1997) Fractionally differenced ARIMA models applied to
hydrologic time series: identification, estimation, and simulation. Water Resour Res 33
(5):1035–1044
Moon Y-I, Lall U (1994) A kernel quantile function estimation for flood frequency analysis. Water
Resour Res 30(11):3095–3103
Moran PAP (1954) A probability theory of dams and storage systems. Aust J Appl Sci 5:116–124
Moss ME, Bryson MC (1974) Autocorrelation structure of monthly streamflows. Water Resour
Res 10(4):737–744
Murrone F, Rossi F, Claps P (1997) Conceptually-based shot noise modeling of streamflows at
short time interval. Stoch Hydrol Hydraul 11(6):483–510
Nagesh Kumar D, Lall L, Peterson MR (2000) Multisite disaggregation of monthly to daily
streamflow. Water Resour Res 36(7):1823–1833
Neyman J, Scott EL (1958) Statistical approach to problems of cosmology. J R Stat Soc Ser B
20:1–29
References 105

Noguchi K, Gel YR, Duguay CR (2011) Bootstrap-based tests for trends in hydrological time
series, with application to ice phenology data. J Hydrol 410:150–161
Northrop P (1998) A clustered spatial-temporal model of rainfall. Proc R Soc London Ser A
454:1875–1888
Nowak K, Prairie J, Rajagopalan B, Lall U (2010) A non-parametric stochastic approach for
multisite disaggregation of annual to daily streamflow. Water Resour Res 46:W08529. doi:10.
1029/2009WR008530
Obeysekera JTB, Tabios G, Salas JD (1987) On parameter estimation of temporal rainfall models.
Water Resour Res 23(10):1837–1850
Olsson J (1998) Evaluation of a scaling cascade model for temporal rainfall disaggregation. Hydrol
Earth Syst Sci 2(1):19–30
Olsson J, Niemczynowicz J, Berndtsson R (1993) Fractal analysis of high-resolution rainfall time
series. J Geophys Res 98(D12):23265–23274
Onof C, Wheater HS (1993) Modelling of British rainfall using a random parameter Bartlett-Lewis
rectangular pulse model. J Hydrol 149:67–95
Onof C, Wheater H (1994) Improvements to the modelling of British rainfall using a modified
random parameter Bartlett-Lewis rectangular pulse model. J Hydrol 157:177–195
Onof C, Chandler RE, Kakou A, Northrop P, Wheater HS, Isham V (2000) Rainfall modelling
using Poisson-cluster processes: a review of developments. Stoch Environ Res Risk Assess 14
(6):384–411
Önöz B, Bayazit M (2012) Block bootstrap for Mann-Kendall trend test of serially dependent data.
Hydrol Process 26(23):3552–3560
Ouarda TBMJ, Labadie JW, Fontane DG (1997) Indexed sequential hydrologic modeling for
hydropower capacity estimation. J Am Water Resour Assoc 33(6):1337–1349
Over TM, Gupta VK (1996) Statistical analysis of mesoscale rainfall: dependence of a random
cascade generator on large-scale forcing. J Appl Meteorol 33:1526–1542
Parzen E (1962) On estimation of a probability density function and mode. Ann Math Statist
33:1065–1076
Pegram GGS, James W (1972) Multilag multivariate autoregressive model for the generation of
operational hydrology. Water Resour Res 8(4):1074–1076
Pelletier JD, Turcotte DL (1997) Long-range persistence in climatological and hydrological time
series: analysis, modeling and application to drought hazard assessment. J Hydrol 203:198–208
Perica S, Foufoula-Georgiou E (1996) Model for multiscale disaggregation of spatial rainfall based
on coupling meteorological and scaling descriptions. J Geophys Res 101:26347–26361
Pham MT, Vanhaute WJ, Vandenberghe S, De Baets B, Verhoest NEC (2013) An assessment of
the ability of Bartlett-Lewis type of rainfall models to reproduce drought statistics. Hydrol
Earth Syst Sci 17(12):5167–5183
Prairie JR, Rajagopalan B, Fulp TJ, Zagona EA (2006) Modified K-NN model for stochastic
streamflow simulation. J Hydrol Eng 11(4):371–378
Prairie J, Rajagopalan R, Lall U, Fulp T (2007) A stochastic nonparametric technique for
space-time disaggregation of streamflows. Water Resour Res 43:W03432. doi:10.1029/
2005WR004721
Puente CE, Bierkens MFP, Diaz-Granados MA, Dik PE, Lopez MM (1993) Practical use of
analytically derived runoff models based on rainfall point processes. Water Resour Res
29:3551–3560
Pui A, Sharma A, Mehrotra R, Sivakumar B, Jeremiah E (2012) A comparison of alternatives for
daily to sub-daily rainfall disaggregation. J Hydrol 470–471:138–157
Quimpo RG (1967) Stochastic model of daily river flow sequences. Hydrology paper 18, Colorado
State University, Fort Collins, Colorado
Rajagopalan B, Lall U (1995) A kernel estimator for discrete distributions. J Nonparametric Statist
4:406–426
Rajagopalan B, Lall U (1999) A k-nearest-neighbor simulator for daily precipitation and other
weather variables. Water Resour Res 35(10):3089–3101
106 3 Stochastic Time Series Methods

Rajagopalan B, Lall U, Tarboton DG (1996) Nonhomogeneous Markov model for daily

precipitation. J Hydrol Eng 1(1):33–40
Rajagopalan B, Lall U, Tarboton DG (1997) Evaluation of kernel density estimation methods for
daily precipitation resampling. Stoch Hydrol Hydraul 11:523–547
Ramirez J, Bras RL (1985) Conditional distributions of Neyman-Scott models for storm arrivals
and their use in irrigation scheduling. Water Resour Res 21(3):317–330
Rao SG, Rao RA (1984) Nonlinear stochastic model of rainfall runoff process. Water Resour Res
20(2):297–309
Rasmussen PF, Salas JD, Fagherazzi L, Bobée B (1996) Estimation and validation of
contemporaneous PARMA models for streamflow simulation. Water Resour Res 32
(10):3151–3160
Regonda SK, Rajagopalan B, Lall U, Clark M, Moon Y-I (2005) Local polynomial method for
ensemble forecast of time series. Nonlinear Proc Geophys 12:397–406
Regonda SK, Rajagopalan B, Clark M, Zagona E (2006a) A multimodel ensemble forecast
framework: Application to spring seasonal flows in the Gunnison River Basin. Water Resour
Res 42:W09404. doi:10.1029/2005WR004653
Regonda SK, Rajagopalan B, Clark M (2006b) A new method to produce categorical streamflow
forecasts. Water Resour Res 42:W09501. doi:10.1029/2006WR004984
Robertson A, Kirshner S, Smyth P (2004) Downscaling of daily rainfall occurrence over northeast
Brazil using a hidden Markov model. J Climate 17(22):4407–4424
Rodriguez-Iturbe I, Rinaldo A (1997) Fractal river basins: chance and self-organization.
Cambridge University Press, Cambridge
Rodriguez-Iturbe I, Gupta VK, Waymire E (1984) Scale considerations in the modeling of
temporal rainfall. Water Resour Res 20(11):1611–1619
Rodriguez-Iturbe I, Cox DR, Eagleson PS (1986) Spatial modeling of total storm rainfall. Proc R
Soc Lond A 403:27–50
Rodriguez-Iturbe I, Cox DR, Isham V (1987) Some models for rainfall based on stochastic point
processes. Proc R Soc London Ser A 410:269–288
Rodriguez-Iturbe I, Cox DR, Isham V (1988) A point process model for rainfall: further
developments. Proc R Soc London Ser A 417:283–298
Roesner LA, Yevjevich VM (1966) Mathematical models for time series of monthly precipitation
and monthly runoff. Colo State Univ Hydol Pap 15
Roldan J, Woolhiser DA (1982) Stochastic daily precipitation models: 1. A comparison of
occurrence processes. Water Resour Res 18:1451–1459
Ross (1996) Stochastic processes. John Wiley & Sons, New York
Sain SR, Baggerly KA, Scott DW (1994) Cross-validation of multivariate densities. J Am Stat
Assoc 89:807–817
Salas JD (1993) Analysis and modeling of hydrologic time series. In: Maidment DR
(ed) Handbook of hydrology. McGraw-Hill, New York, pp 19.1–19.72
Salas JD, Abdelmohsen MW (1993) Initialization for generating single-site and multisite
low-order periodic autoregressive and moving average processes. Water Resour Res 29
(6):1771–1776
Salas JD, Lee T (2010) Nonparametric simulation of single site seasonal streamflow. J Hydrol Eng
15(4):284–296
Salas JD, Obeysekera JTB (1982) ARMA model identification of hydrologic time series. Water
Resour Res 18(4):1011–1021
Salas JD, Obeysekera JTB (1992) Conceptual basis of seasonal streamflow time series models.
J Hydraul Eng 118(8):1186–1194
Salas JD, Smith RA (1981) Physical basis of stochastic models of annual flows. Water Resour Res
17(2):428–430
Salas JD, Boes DC, Yevjevich V, Pegram GGS (1979) Hurst phenomenon as a pre-asymptotic
behavior. J Hydrol 44(1):1–15
Salas JD, Boes DC, Smith RA (1982) Estimation of ARMA models with seasonal parameters.
Water Resour Res 18(4):1006–1010
References 107

Salas JD, Delleur JR, Yevjevich VM, Lane WL (1995) Applied modeling of hydrologic time
series. Water Resources Publications, Littleton, Colorado
Santos EG, Salas JD (1992) Stepwise disaggregation scheme for synthetic hydrology. J Hydraul
Eng 118(5):765–784
Schertzer D, Lovejoy S (1987) Physical modeling and analysis of rain and clouds by anisotropic
scaling multiplicative processes. J Geophys Res 92(D8):9693–9714
Schuster E, Yakowitz S (1985) Parametric/nonparametric mixture density estimation with
application to flood-frequency analysis. Water Resour Bull 21(5):797–803
Scott DW (1992) Multivariate density estimation: theory, practice and visualization. John Wiley
and Sons Inc, New York
Seed AW, Draper C, Srikanthan R, Menabde M (2000) A multiplicative broken-line model for
time series of mean areal rainfall. Water Resour Res 36(8):2395–2399
Şen Z (1976) Wet and dry periods of annual flow series. J Hydraul Div Amer Soc Civ Eng
102:1503–1514
Şen Z (1990) Critical drought analysis by second-order Markov chain. J Hydrol 120:183–202
Shahjahan Mondal S, Wasimi SA (2006) Generating and forecasting monthly flows of the Ganges
river with PAR model. J Hydrol 323:41–56
Sharifazari S, Araghinejad S (2015) Development of a nonparametric model for multivariate
hydrological monthly series simulation considering climate change impacts. Water Resour
Manage 29(14):5309–5322
Sharma A, O’Neill R (2002) A nonparametric approach for representing interannual dependence in
monthly streamflow sequences. Water Resour Res 38(7):1100. doi:10.1029/2001WR000953
Sharma A, Tarboton DG, Lall U (1997) Streamflow simulation: a nonparametric approach. Water
Resour Res 33(2):291–308
Sharma A, Lall U, Tarboton DG (1998) Kernel bandwidth selection for a first order nonparametric
streamflow simulation model. Stoch Hydrol Hydraul 12:33–52
Sharma TC, Panu US (2014) A simplified model for predicting drought magnitudes: a case of
streamflow droughts in Canadian Prairies. Water Resour Manage 28:1597–1611
Sikorska AE, Montanari A, Koutsoyiannis D (2015) Estimating the uncertainty of hydrological
predictions through data-driven resampling techniques. J Hydrol Eng 20(1):A4014009
Silverman BW (1986) Density estimation for statistics and data analysis. Chapman & Hall, New
York
Singhrattna N, Rajagopalan B, Clark M, Krishna Kumar K (2005) Forecasting Thailand summer
monsoon rainfall. Int J Climatol 25(5):649–664
Sivakumar B (2000) Fractal analysis of rainfall observed in two different climatic regions. Hydrol
Sci J 45(5):727–738
Smith JA (1991) Long-range streamflow forecasting using nonparametric regression. Water
Resour Bull 27(1):39–46
Smith JA, Carr AF (1983) A point process model of summer season rainfall occurrences. Water
Resour Res 19(1):95–103
Smith JA, Day GN, Kane MD (1992) Nonparametric framework for long range streamflow
forecasting. J Water Res Plng Mgmt 118(1):82–92
Smithers JC, Pegram GGS, Schulze RE (2002) Design rainfall estimation in South Africa using
Bartlett-Lewis rectangular pulse rainfall models. J Hydrol 258(1–4):83–99
Sonali P, Kumar Nagesh (2013) Review of trend detection methods and their application to detect
temperature changes In India. J Hydrol 476:212–227
Souza Filho FA, Lall U (2003) Seasonal to interannual ensemble streamflow forecasts for Ceara,
Brazil: Applications of a multivariate semiparametric algorithm. Water Resour Res 39
(11):1307. doi:10.1029/2002WR001373
Spitzer F (1964) Principles of random walk. Springer, New York
Sposito G (2008) Scale dependence and scale invariance in hydrology. Cambridge University
Press, Cambridge, UK
Sprent P, Smeeton NC (2007) Applied nonparametric statistical methods. Chapman & Hall, Boca
Raton
108 3 Stochastic Time Series Methods

Srikanthan R, McMahon TA (1983) Stochastic simulation of daily rainfall for Australian stations.
Trans ASAE 26:754–759
Srikanthan R, Harrold TI, Sharma A, McMahon TA (2005) Comparison of two approaches for
generation of daily rainfall data. Stoch Environ Res Risk Assess 19:215–226
Srinivas VV, Srinivasan K (2001) A hybrid stochastic model for multiseason streamflow
simulation. Water Resour Res 37(10):2537–2549
Srinivas VV, Srinivasan K (2005) Hybrid moving block bootstrap for stochastic simulation of
multi-site multi-season streamflows. J Hydrol 302(1–4):307–330
Srinivas VV, Srinivasan K (2006) Hybrid matched-block bootstrap for stochastic simulation of
multiseason streamflows. J Hydrol 329(1–2):1–15
Srivatsav RK, Simonovic SP (2014) Multi-site, multivariate weather generator using maximum
entropy bootstrap. Clim Dyn 44:3431–3448
Srivatsav RK, Srinivasan K, Sudheer KP (2011) Simulation-optimization framework for
multi-season hybrid stochastic models. J Hydrol 404(3):209–225
Stagge JH, Moglen GE (2013) A nonparametric stochastic method for generating daily
climate-adjusted streamflows. Water Resour Res 49:6179–6193
Stedinger JR, Taylor MR (1982) Synthetic streamflow generation: 1. Model verification and
validation. Water Resour Res 18(4):909–918
Stedinger JR, Vogel RM (1984) Disaggregation procedures for generating serially correlated flow
vectors. Water Resour Res 20(1):47–56
Stedinger JR, Lettenmaier DP, Vogel RM (1985) Multisite ARMA(1,1) and disaggregation models
for annual streamflow generation. Water Resour Res 21(4):497–509
Tao PC, Delluer JW (1976) Multistation, multiyear synthesis of hydrologic time series by
disaggregation. Water Resour Res 12(6):1303–1312
Tarboton DG, Sharma A, Lall U (1998) Disaggregation procedures for stochastic hydrology based
on nonparametric density estimation. Water Resour Res 34(1):107–119
Tesfaye YG, Meerschaert MM, Anderson PL (2006) Identification of periodic autoregressive
moving average models and their application to the modeling of river flows. Water Resour Res
42:W01419. doi:10.1029/2004WR003772
Tessier Y, Lovejoy S, Hubert P, Schertzer D, Pecknold S (1996) Multifractal analysis and
modeling of rainfall and river flows and scaling, causal transfer functions. J Geophys Res 101
(D21):26427–26440
Thomas HA, Fiering MB (1962) Mathematical synthesis of streamflow sequences for the analysis
of river basins by simulation. In: Mass A et al (eds) Design of water resource systems. Harvard
University Press, Cambridge, Massachusetts, pp 459–493
Thyer MA, Kuczera G (2000) Modelling long-term persistence in hydro-climatic time series using
a hidden state Markov model. Water Resour Res 36:3301–3310
Todini E (1980) The preservation of skewness in linear disaggregation schemes. J Hydrol 47:199–
214
Towler E, Rajagopalan B, Seidel C, Summers RS (2009) Simulating ensembles of source water
quality using a k-nearest neighbor resampling approach. Environ Sci Technol 43:1407–1411
Ula TA (1990) Periodic covariance stationarity of multivariate periodic autoregressive moving
average processes. Water Resour Res 26(5):855–861
Valencia DR, Schaake JC (1973) Disaggregation processes in stochastic hydrology. Water Resour
Res 9(3):580–585
Vecchia AV (1985) Periodic autoregressive-moving average (PARMA) modeling with applica-
tions to water resources. Water Resour Bull 21(5):721–730
Veneziano D, Langousis A (2010) Scaling and fractals in hydrology. In: Sivakumar B,
Berndsson R (eds) Advances in data-based approaches for hydrologic modeling and
forecasting. World Scientific, Singapore, pp 107–243
Veneziano D, Moglen G, Furcolo P, Iacobellis V (2000) Stochastic model of the width function.
Water Resour Res 36(4):1143–1157
Veneziano D, Langousis A, Furcolo P (2006) Multifractality and rainfall extremes: A review.
Water Resour Res 42:W06D15. doi:10.1029/2005WR004716
References 109

Venugopal V, Foufoula-Georgiou E, Sapozhnikov V (1999) A space-time downscaling model for

rainfall. J Geophys Res-Atmos 104(D16):19705–19721
Verdin A, Funk C, Rajagopalan B, Kleiber W (2016) Kriging and local polynomial methods for
blending satellite-derived and gauge precipitation estimates to support hydrologic early
warning systems. IEEE Trans Geoscience and Remote Sensing 54(5):2552–2562
Verhoest N, Troch PA, De Troch FP (1997) On the applicability of Bartlett-Lewis rectangular
pulses models in the modeling of design storms at a point. J Hydrol 202(1–4):108–120
Verhoest NEC, Vandenberghe S, Cabus P, Onof C, Meca-Figueras T, Jameleddine S (2010) Are
stochastic point rainfall models able to preserve extreme flood statistics? Hydrol Process
24:3439–3445
Viola F, Francipane A, Caracciolo D, Pumo D, La Loggia G, Noto LV (2016) Co-evolution of
hydrological components under climate change scenarios in the Mediterranean area. Sci Total
Environ 544:515–524
Vogel RM, Shallcross AL (1996) The moving blocks bootstrap versus parametric time series
models. Water Resour Res 32(6):1875–1882
Vrugt JA (2016) Markov chain Monte Carlo simulation using the DREAM software package:
theory, concepts, and MATLAB implementation. Environ Modell Softw 75:273–316
Vrugt JA, ter Braak CJF, Clark MP, Hyman JM, Robinson BA (2008) Treatment of input
uncertainty in hydrologic modeling: doing hydrology backward with Markov chain Monte
Carlo simulation. Water Resour Res 44:W00B09. doi:10.1029/2007WR006720
Wasko C, Pui A, Sharma A, Mehrotra R, Jeremiah E (2015) Representing low-frequency
variability in continuous rainfall simulations: a hierarchical random Bartlett Lewis continuous
rainfall generation model. Water Resour Res 51(12):9995–10007
Waymire E, Gupta VK, Rodriguez-Iturbe I (1984) A spectral theory of rainfall intensity at the
meso-β scale. Water Resour Res 20(10):1453–1465
Wiener N (1949) Extrapolation, interpolation, and smoothing of stationary time series. MIT Press
Wilks DL (1998) Multisite generalisation of a daily stochastic precipitation generation model.
J Hydrol 210:178–191
Wilks DS (2010) Use of stochastic weather generators for precipitation downscaling. Wiley
Interdiscip Rev 1:809–907
Wong H, Ip W-C, Zhang R, Xia J (2007) Nonparametric time series models for hydrologic
forecasting. J Hydrol 332:337–347
Woolhiser DA, Osborne HB (1985) A stochastic model of dimensionless thunderstorm rainfall.
Water Resour Res 21(4):511–522
Yakowitz S (1973) A stochastic model for daily river flows in an arid region. Water Resour Res 9
(5):1271–1285
Yakowitz S (1979) A nonparametric Markov model for daily river flow. Water Resour Res 15
(5):1035–1043
Yakowitz S (1985) Markov flow models and the flood warning problem. Water Resour Res 21
(1):81–88
Yakowitz (1993) Nearest-neighbor regression estimation for null-recurrent Markov time series.
Stochastic Processes Their Appl 48:311–318
Yates D, Gangopadhyay S, Rajagopalan B, Strzepek K (2003) A technique for generating regional
climate scenarios using a nearest neighbor bootstrap. Water Resour Res 39(7):1199. doi:10.
1029/2002WR001769
Yevjevich VM (1963) Fluctuations of wet and dry years—Part I, Research data assembly and
mathematical models. Hydrology paper 1, Colorado State University, Fort Collins, Colorado
Yevjevich VM (1964) Fluctuations of wet and dry years—Part II, analysis by serial correlation.
Hydrology paper 4, Colorado State University, Fort Collins, Colorado
Yevjevich VM (1972) Stochastic processes in hydrology. Water Resour Publ, Fort Collins,
Colorado
Yevjevich VM (1975) Generation of hydrologic time series. Hydrology paper 56, Colorado State
University, Fort Collins, CO
110 3 Stochastic Time Series Methods

Yue S, Pilon P (2004) A comparison of the power of the t test, Mann-Kendall and bootstrap tests
for trend detection. Hydrol Sci J 49(1):21–37
Zucchini W, Adamson PT (1989) Bootstrap confidence intervals for design storms from
exceedence series. Hydrol Sci J 34:41–48
Zucchini W, Guttorp P (1991) A hidden Markov model for space-time precipitation. Water Resour
Res 27:1917–1923